IBM XL Fortran for Blue Gene/Q, V14.1
Optimization and Programming Guide
Ve r s i o n 1 4 .1
SC14-7370-00
IBM XL Fortran for Blue Gene/Q, V14.1
Optimization and Programming Guide
Ve r s i o n 1 4 .1
SC14-7370-00 Note Before using this information and the product it supports, read the information in “Notices” on page 311.
First edition This edition applies to IBM XL Fortran for Blue Gene/Q, V14.1 (Program 5799-AH1) and to all subsequent releases and modifications until otherwise indicated in new editions. Make sure you are using the correct edition for the level of the product. © Copyright IBM Corporation 1990, 2012. US Government Users Restricted Rights – Use, duplication or disclosure restricted by GSA ADP Schedule Contract with IBM Corp. Contents
About this information ...... vii Chapter 3. Advanced optimization Who should read this information...... vii concepts ...... 41 How to use this information ...... vii Aliasing ...... 41 How this information is organized ...... vii Inlining...... 41 Conventions ...... viii Finding the right level of inlining ...... 42 Related information ...... xii IBM XL Fortran information...... xii Chapter 4. Managing code size ....45 Standards and specifications ...... xiii Steps for reducing code size ...... 46 Other IBM information ...... xiv Compiler option influences on code size.....46 Technical support ...... xiv The -qipa compiler option ...... 46 How to send your comments ...... xiv The -qinline inlining option ...... 46 The -qhot compiler option ...... 47 Chapter 1. Optimizing your applications 1 The -qcompact compiler option...... 47 Distinguishing between optimization and tuning . . 1 Other influences on code size ...... 47 Steps in the optimization process ...... 2 High activity areas ...... 47 Basic optimization ...... 2 Computed GOTOs and CASE constructs . . . 47 Optimizing at level 0 ...... 3 Code size with dynamic or static linking . . . 48 Optimizing at level 2 ...... 3 Advanced optimization ...... 4 Chapter 5. Compiler-friendly Optimizing at level 3 ...... 5 programming techniques ...... 51 An intermediate step: adding -qhot suboptions at level 3 ...... 6 General practices ...... 51 Optimizing at level 4 ...... 7 Variables and pointers ...... 51 Optimizing at level 5 ...... 8 Arrays ...... 52 Specialized optimization techniques ...... 8 Choosing appropriate variable sizes ...... 52 High-order transformation (HOT) ...... 9 Interprocedural analysis (IPA) ...... 11 Chapter 6. High performance libraries 55 Vector technology ...... 14 Using the Mathematical Acceleration Subsystem Using compiler reports to diagnose optimization libraries (MASS) ...... 55 opportunities ...... 18 Using the scalar library ...... 56 Debugging optimized code ...... 20 Using the vector libraries ...... 58 Understanding different results in optimized Using the SIMD library ...... 62 programs ...... 21 Compiling and linking a program with MASS . . 64 Debugging in the presence of optimization . . . 21 Using the Basic Linear Algebra Subprograms – Using -qoptdebug to help debug optimized BLAS ...... 65 programs ...... 22 BLAS function syntax ...... 65 Tracing procedures in your code ...... 25 Linking the libxlopt library ...... 67 Getting more performance ...... 29 Beyond performance: effective programming Chapter 7. Parallel programming with techniques ...... 30 XL Fortran ...... 69 Thread-level speculative execution...... 69 Chapter 2. Tuning XL compiler Transactional memory ...... 70 applications ...... 31 Profiler for OpenMP ...... 72 Tuning for your target architecture ...... 31 Compiling your parallelized code ...... 72 Using -qarch ...... 31 The _OPENMP C preprocessor macro and Using -qtune ...... 32 conditional compilation ...... 73 Using -qcache ...... 33 Setting runtime options ...... 74 Before you finish tuning ...... 33 XLSMPOPTS ...... 74 Tuning your code for Blue Gene ...... 33 BG_SMP_FAST_WAKEUP ...... 78 Further option driven tuning ...... 35 BG_SPEC_SCRUB_CYCLE ...... 79 Options for providing application characteristics 36 Environment variables for OpenMP .....79 Options to control optimization transformations 38 Environment variables for thread-level Options to assist with performance analysis . . 39 speculative execution ...... 85 Options that can inhibit performance .....40 Environment variables for transactional memory 86 Optimizing your SMP code ...... 89
© Copyright IBM Corp. 1990, 2012 iii Developing and running SMP applications . . . 89 f_pthread_mutex_t ...... 233 Parallelization directives ...... 90 f_pthread_mutex_trylock(mutex) ...... 233 An introduction to parallelization directives . . 90 f_pthread_mutex_unlock(mutex) ...... 234 Detailed descriptions of parallelization directives 93 f_pthread_mutexattr_destroy(mattr) .....234 Directive clauses ...... 146 f_pthread_mutexattr_getpshared(mattr, pshared) 235 Routines for parallel programming ...... 173 f_pthread_mutexattr_gettype(mattr, type) . . . 236 Routines for OpenMP ...... 173 f_pthread_mutexattr_init(mattr) ...... 237 Routines for thread-level speculative execution 195 f_pthread_mutexattr_setpshared(mattr, pshared) 237 Routines for transactional memory .....197 f_pthread_mutexattr_settype(mattr, type) . . . 238 Pthreads library module...... 201 f_pthread_mutexattr_t ...... 239 Pthreads data structures, functions, and f_pthread_once(once, initr) ...... 239 subroutines ...... 202 f_pthread_once_t ...... 239 f_maketime(delay)...... 204 f_pthread_rwlock_destroy(rwlock) .....240 f_pthread_attr_destroy(attr)...... 205 f_pthread_rwlock_init(rwlock, rwattr) ....240 f_pthread_attr_getdetachstate(attr, detach) . . . 205 f_pthread_rwlock_rdlock(rwlock) ...... 241 f_pthread_attr_getguardsize(attr, guardsize) . . 206 f_pthread_rwlock_t ...... 241 f_pthread_attr_getinheritsched(attr, inherit) . . 206 f_pthread_rwlock_tryrdlock(rwlock) .....242 f_pthread_attr_getschedparam(attr, param) . . 207 f_pthread_rwlock_trywrlock(rwlock) ....242 f_pthread_attr_getschedpolicy(attr, policy) . . . 208 f_pthread_rwlock_unlock(rwlock) .....243 f_pthread_attr_getscope(attr, scope) .....208 f_pthread_rwlock_wrlock(rwlock) .....243 f_pthread_attr_getstack(attr, stackaddr, ssize) 209 f_pthread_rwlockattr_destroy(rwattr) ....244 f_pthread_attr_init(attr) ...... 209 f_pthread_rwlockattr_getpshared(rwattr, f_pthread_attr_setdetachstate(attr, detach) . . . 210 pshared) ...... 244 f_pthread_attr_setguardsize(attr, guardsize) . . 211 f_pthread_rwlockattr_init(rwattr) ...... 245 f_pthread_attr_setinheritsched(attr, inherit) . . 211 f_pthread_rwlockattr_setpshared(rwattr, f_pthread_attr_setschedparam(attr, param) . . 212 pshared) ...... 246 f_pthread_attr_setschedpolicy(attr, policy) . . . 213 f_pthread_rwlockattr_t ...... 246 f_pthread_attr_setscope(attr, scope) .....213 f_pthread_self()...... 247 f_pthread_attr_setstack(attr, stackaddr, ssize) 214 f_pthread_setcancelstate(state, oldstate)....247 f_pthread_attr_t ...... 215 f_pthread_setcanceltype(type, oldtype) ....248 f_pthread_cancel(thread) ...... 215 f_pthread_setconcurrency(new_level) ....249 f_pthread_cleanup_pop(exec) ...... 215 f_pthread_setschedparam(thread, policy, param) 249 f_pthread_cleanup_push(cleanup, flag, arg) . . 216 f_pthread_setspecific(key, arg) ...... 250 f_pthread_cond_broadcast(cond) ...... 217 f_pthread_t ...... 251 f_pthread_cond_destroy(cond)...... 218 f_pthread_testcancel() ...... 251 f_pthread_cond_init(cond, cattr) ...... 218 f_sched_param ...... 251 f_pthread_cond_signal(cond) ...... 219 f_sched_yield() ...... 251 f_pthread_cond_t ...... 219 f_timespec ...... 252 f_pthread_cond_timedwait(cond, mutex, timeout) ...... 220 Chapter 8. Interlanguage calls ....253 f_pthread_cond_wait(cond, mutex) .....220 Conventions for XL Fortran external names . . . 253 f_pthread_condattr_destroy(cattr)...... 221 Mixed-language input and output ...... 254 f_pthread_condattr_getpshared(cattr, pshared) 221 Mixing Fortran and C++ ...... 254 f_pthread_condattr_init(cattr) ...... 222 Making calls to C functions work...... 256 f_pthread_condattr_setpshared(cattr, pshared) 223 Passing data from one language to another . . . 257 f_pthread_condattr_t ...... 223 Passing arguments between languages ....257 f_pthread_create(thread, attr, flag, ent, arg) . . 224 Passing global variables between languages . . 258 f_pthread_detach(thread) ...... 225 Passing character types between languages . . 259 f_pthread_equal(thread1, thread2) .....226 Passing arrays between languages .....260 f_pthread_exit(ret) ...... 226 Passing pointers between languages .....261 f_pthread_getconcurrency() ...... 227 Passing arguments by reference or by value . . 261 f_pthread_getschedparam(thread, policy, param) 227 Passing COMPLEX values to/from gcc ....263 f_pthread_getspecific(key, arg)...... 228 Returning values from Fortran functions . . . 263 f_pthread_join(thread, ret) ...... 229 Arguments with the OPTIONAL attribute . . . 263 f_pthread_key_create(key, dtr) ...... 229 Assembler-level subroutine linkage conventions 264 f_pthread_key_delete(key) ...... 230 The stack ...... 265 f_pthread_key_t ...... 231 The Linkage Area and Minimum Stack Frame 266 f_pthread_kill(thread, sig) ...... 231 The input parameter area ...... 267 f_pthread_mutex_destroy(mutex) ...... 231 The register save area ...... 267 f_pthread_mutex_init(mutex, mattr) .....232 The local stack area ...... 267 f_pthread_mutex_lock(mutex) ...... 233 The output parameter area ...... 267 iv XL Fortran: Optimization and Programming Guide Linkage convention for argument passing ....267 Minimizing rounding errors ...... 292 Argument passing rules (by value) .....268 Minimizing overall rounding ...... 292 Order of arguments in argument list ....269 Delaying rounding until run time .....292 Linkage convention for function calls .....269 Ensuring that the rounding mode is consistent 292 Pointers to functions ...... 270 Duplicating the floating-point results of other Function values ...... 270 systems ...... 293 The stack floor ...... 271 Maximizing floating-point performance ....293 Stack overflow ...... 271 Detecting and trapping floating-point exceptions 294 Prolog and epilog ...... 271 Compiler features for trapping floating-point Traceback...... 272 exceptions ...... 294 Installing an exception handler ...... 295 Chapter 9. Implementation details of Producing a core file ...... 296 XL Fortran Input/Output (I/O).....273 Controlling the floating-point status and control register ...... 296 Implementation details of file formats .....273 xlf_fp_util procedures ...... 297 File names ...... 274 fpgets and fpsets subroutines ...... 297 Preconnected and Implicitly Connected Files . . . 275 Sample programs for exception handling . . . 299 File positioning...... 276 Causing exceptions for particular variables . . 299 I/O redirection ...... 276 Minimizing the performance impact of How XL Fortran I/O interacts with pipes, special floating-point exception trapping ...... 299 files, and links ...... 277 Default record lengths ...... 277 File permissions ...... 277 Chapter 11. Porting programs to XL Selecting error messages and recovery actions . . 278 Fortran ...... 301 Flushing I/O buffers ...... 278 Outline of the porting process ...... 301 Choosing locations and names for Input/Output Portability of directives ...... 301 files ...... 279 Common industry extensions that XL Fortran Naming files that are connected with no explicit supports ...... 302 name ...... 279 Mixing data types in statements ...... 303 Naming scratch files ...... 279 Date and time routines ...... 304 XL Fortran thread-safe I/O library ...... 280 Other libc routines ...... 304 Synchronization of I/O operations .....280 Changing the default sizes of data types . . . 304 Parallel I/O issues...... 281 Name conflicts between your procedures and Use of I/O statements in signal handlers . . . 283 XL Fortran intrinsic procedures ...... 304 Asynchronous thread cancellation ...... 283 Reproducing results from other systems . . . 304
Chapter 10. Implementation details of Chapter 12. Sample Fortran programs 305 XL Fortran floating-point processing . 285 Example1-XLFortran source file ...... 305 IEEE floating-point overview ...... 285 Example 2 - valid C routine source file .....305 Compiling for strict IEEE conformance ....285 Example 3 - valid Fortran SMP source file ....308 IEEE Single- and double-precision values . . . 286 Example 4 - invalid Fortran SMP source file . . . 308 IEEE extended-precision values ...... 286 Programming examples using the Pthreads library Infinities and NaNs ...... 286 module ...... 309 Exception-handling model ...... 287 Hardware-specific floating-point overview....288 Notices ...... 311 Single- and double-precision values .....288 Trademarks and service marks ...... 313 Extended-precision values ...... 289 How XL Fortran rounds floating-point calculations 290 Index ...... 315 Selecting the rounding mode ...... 290
Contents v vi XL Fortran: Optimization and Programming Guide About this information
This information is part of the IBM® XL Fortran for Blue Gene®/Q, V14.1 information suite. It provides both reference information and practical tips for using XL Fortran's optimization and tuning capabilities to maximize application performance, as well as expanding on programming concepts such as I/O and interlanguage calls.
Who should read this information This information is for anyone who wants to exploit the XL Fortran compiler's capabilities for optimizing and tuning Fortran programs. Readers should be familiar with their Blue Gene/Q operating system and have extensive Fortran programming experience with complex applications. However, users new to XL Fortran can still use this information to help them understand how the compiler's features can be used for effective program optimization.
How to use this information This guide focuses on specific programming and compilation techniques that can maximize XL Fortran application performance. It covers optimization and tuning strategies, recommended programming practices and compilation procedures, debugging, and information on using XL Fortran advanced language features. This guide also contains cross-references to relevant topics of other reference guides in the XL Fortran information suite.
Topics not described in this information are available as indicated in the following: v Installation, system requirements, last-minute updates: see the XL Fortran Installation Guide and product README. v Overview of XL Fortran features: see the Getting Started with XL Fortran. v Syntax, semantics, and implementation of the XL Fortran programming language: see the XL Fortran Language Reference. v Compiler setup, compiling and running programs, compiler options, diagnostics: see the XL Fortran Compiler Reference. v Operating system commands related to the use of the compiler: consult your Linux-specific distribution's man page help and information.
How this information is organized This guide includes the following topics: v Chapter 1, “Optimizing your applications,” on page 1 provides an overview of the optimization process. v Chapter 2, “Tuning XL compiler applications,” on page 31 discusses the compiler options available for optimizing and tuning code. v Chapter 3, “Advanced optimization concepts,” on page 41, Chapter 4, “Managing code size,” on page 45, and “Debugging optimized code” on page 20 discuss advanced techniques like optimizing loops and inlining code, and debug considerations for optimized code.
© Copyright IBM Corp. 1990, 2012 vii v The following sections contain information on how to write optimization friendly, portable XL Fortran code, that is interoperable with other languages. Also included is a description of XL Fortran's OpenMP and SMP support with guidelines for writing parallel code. – Chapter 5, “Compiler-friendly programming techniques,” on page 51 – Chapter 6, “High performance libraries,” on page 55 – Chapter 7, “Parallel programming with XL Fortran,” on page 69 – Chapter 8, “Interlanguage calls,” on page 253 v The following sections contain information about XL Fortran and its implementation that can be useful for new and experienced users alike, as well as those who want to move their existing Fortran applications to the XL Fortran compiler: – Chapter 9, “Implementation details of XL Fortran Input/Output (I/O),” on page 273 – Chapter 10, “Implementation details of XL Fortran floating-point processing,” on page 285 – Chapter 11, “Porting programs to XL Fortran,” on page 301
Conventions Typographical conventions
The following table shows the typographical conventions used in the IBM XL Fortran for Blue Gene/Q, V14.1 information. Table 1. Typographical conventions Typeface Indicates Example bold Lowercase commands, executable The compiler provides basic names, compiler options, and invocation commands, bgxlf, along directives. with several other compiler invocation commands to support various Fortran language levels and compilation environments. italics Parameters or variables whose Make sure that you update the size actual names or values are to be parameter if you return more than supplied by the user. Italics are the size requested. also used to introduce new terms. underlining The default setting of a parameter nomaf | maf of a compiler option or directive. monospace Programming keywords and To compile and optimize library functions, compiler builtins, myprogram.f, enter: bgxlf examples of program code, myprogram.f -O3. command strings, or user-defined names. UPPERCASE Fortran programming keywords, The ASSERT directive applies only to bold statements, directives, and intrinsic the DO loop immediately following procedures. Uppercase letters may the directive, and not to any nested also be used to indicate the DO loops. minimum number of characters required to invoke a compiler option/suboption.
viii XL Fortran: Optimization and Programming Guide Qualifying elements (icons and bracket separators)
In descriptions of language elements, this information uses icons and marked bracket separators to delineate the Fortran language standard text as follows: Table 2. Qualifying elements Bracket Icon separator text Meaning
F2008 N/A The text describes an IBM XL Fortran implementation of the Fortran 2008 standard. F2008
Fortran 2003 The text describes an IBM XL Fortran implementation of begins / ends the Fortran 2003 standard, and it applies to all later standards.
IBM extension The text describes a feature that is an IBM XL Fortran begins / ends extension to the standard language specifications.
Note: If the information is marked with a Fortran language standard icon or bracket separators, it applies to this specific Fortran language standard and all later ones. If it is not marked, it applies to all Fortran language standards. Syntax diagrams
Throughout this information, diagrams illustrate XL Fortran syntax. This section will help you to interpret and use those diagrams. v Read the syntax diagrams from left to right, from top to bottom, following the path of the line. The ─── symbol indicates the beginning of a command, directive, or statement. The ─── symbol indicates that the command, directive, or statement syntax is continued on the next line. The ─── symbol indicates that a command, directive, or statement is continued from the previous line. The ─── symbol indicates the end of a command, directive, or statement. Fragments, which are diagrams of syntactical units other than complete commands, directives, or statements, start with the │─── symbol and end with the ───│ symbol. IBM XL Fortran extensions are marked by a number in the syntax diagram with an explanatory note immediately following the diagram. Program units, procedures, constructs, interface blocks and derived-type definitions consist of several individual statements. For such items, a box encloses the syntax representation, and individual syntax diagrams show the required order for the equivalent Fortran statements. v Required items are shown on the horizontal line (the main path):
keyword required_argument v Optional items are shown below the main path:
About this information ix keyword optional_argument
Note: Optional items (not in syntax diagrams) are enclosed by square brackets ([ and ]). For example, [UNIT=]u v If you can choose from two or more items, they are shown vertically, in a stack. If you must choose one of the items, one item of the stack is shown on the main path.
keyword required_argument1 required_argument2
If choosing one of the items is optional, the entire stack is shown below the main path.
keyword optional_argument1 optional_argument2
v An arrow returning to the left above the main line (a repeat arrow) indicates that you can make more than one choice from the stacked items or repeat an item. The separator character, if it is other than a blank, is also indicated:
,
keyword repeatable_argument
v The item that is the default is shown above the main path.
default_argument keyword alternate_argument
v Keywords are shown in nonitalic letters and should be entered exactly as shown. v Variables are shown in italicized lowercase letters. They represent user-supplied names or values. If a variable or user-specified name ends in _list, you can provide a list of these terms separated by commas. v If punctuation marks, parentheses, arithmetic operators, or other such symbols are shown, you must enter them as part of the syntax.
Sample syntax diagram
The following is an example of a syntax diagram with an interpretation:
x XL Fortran: Optimization and Programming Guide , (1) EXAMPLE char_constant a e name_list b c d
Notes: 1 IBM extension
Interpret the diagram as follows: v Enter the keyword EXAMPLE. v EXAMPLE is an IBM extension. v Enter a value for char_constant. v Enter a value for a or b, but not for both. v Optionally, enter a value for c or d. v Enter at least one value for e. If you enter more than one value, you must put a comma between each. v Enter the value of at least one name for name_list. If you enter more than one value, you must put a comma between each. (The _list syntax is equivalent to the previous syntax for e.)
How to read syntax statements
Syntax statements are read from left to right: v Individual required arguments are shown with no special notation. v When you must make a choice between a set of alternatives, they are enclosed by { and } symbols. v Optional arguments are enclosed by [ and ] symbols. v When you can select from a group of choices, they are separated by | characters. v Arguments that you can repeat are followed by ellipses (...).
Example of a syntax statement EXAMPLE char_constant {a|b}[c|d]e[,e]... name_list{name_list}...
The following list explains the syntax statement: v Enter the keyword EXAMPLE. v Enter a value for char_constant. v Enter a value for a or b, but not for both. v Optionally, enter a value for c or d. v Enter at least one value for e. If you enter more than one value, you must put a comma between each. v Optionally, enter the value of at least one name for name_list. If you enter more than one value, you must put a comma between each name.
Note: The same example is used in both the syntax-statement and syntax-diagram representations.
About this information xi Examples in this information
The examples in this information, except where otherwise noted, are coded in a simple style that does not try to conserve storage, check for errors, achieve fast performance, or demonstrate all possible methods to achieve a specific result.
The examples for installation information are labelled as either Example or Basic example. Basic examples are intended to document a procedure as it would be performed during a basic, or default, installation; these need little or no modification. Notes on the terminology used
Some of the terminology in this information is shortened as follows: v The term free source form format often appears as free source form. v The term fixed source form format often appears as fixed source form. v The term XL Fortran often appears as XLF.
Related information
The following sections provide related information for XL Fortran: IBM XL Fortran information
XL Fortran provides product information in the following formats: v README files README files contain late-breaking information, including changes and corrections to the product information. README files are located by default in the XL Fortran directory and in the root directory of the installation CD. v Installable man pages Man pages are provided for the compiler invocations and all command-line utilities provided with the product. Instructions for installing and accessing the man pages are provided in the IBM XL Fortran for Blue Gene/Q, V14.1 Installation Guide. v Information center The information center of searchable HTML files can be launched on a network and accessed remotely or locally. Instructions for installing and accessing the online information center are provided in the IBM XL Fortran for Blue Gene/Q, V14.1 Installation Guide. The information center of searchable HTML files is viewable on the web at http://pic.dhe.ibm.com/infocenter/compbg/v121v141/index.jsp. v PDF documents PDF documents are located by default in the /opt/ibmcmp/xlf/bg/14.1/doc/ en_US/pdf/ directory. The PDF files are also available on the web at http://www.ibm.com/software/awdtools/fortran/xlfortran/features/bg/ library/. The following files comprise the full set of XL Fortran product information:
xii XL Fortran: Optimization and Programming Guide Table 3. XL Fortran PDF files PDF file Document title name Description IBM XL Fortran for Blue install.pdf Contains information for installing XL Fortran Gene/Q, V14.1 Installation and configuring your environment for basic Guide, GC14-7367-00 compilation and program execution. Getting Started with IBM getstart.pdf Contains an introduction to the XL Fortran XL Fortran for Blue product, with information on setting up and Gene/Q, V14.1, configuring your environment, compiling and GC14-7366-00 linking programs, and troubleshooting compilation errors. IBM XL Fortran for Blue compiler.pdf Contains information about the various Gene/Q, V14.1 Compiler compiler options and environment variables. Reference, GC14-7368-00 IBM XL Fortran for Blue langref.pdf Contains information about the Fortran Gene/Q, V14.1 Language programming language as supported by IBM, Reference, GC14-7369-00 including language extensions for portability and conformance to nonproprietary standards, compiler directives and intrinsic procedures. IBM XL Fortran for Blue proguide.pdf Contains information on advanced Gene/Q, V14.1 Optimization programming topics, such as application and Programming Guide, porting, interlanguage calls, floating-point SC14-7370-00 operations, input/output, application optimization and parallelization, and the XL Fortran high-performance libraries.
To read a PDF file, use the Adobe Reader. If you do not have the Adobe Reader, you can download it (subject to license terms) from the Adobe website at http://www.adobe.com.
More information related to XL Fortran including IBM Redbooks® publications, white papers, tutorials, and other articles, is available on the web at:
http://www.ibm.com/software/awdtools/fortran/xlfortran/features/bg/library/ Standards and specifications
XL Fortran is designed to support the following standards and specifications. You can refer to these standards for precise definitions of some of the features found in this information. v American National Standard Programming Language FORTRAN, ANSI X3.9-1978. v American National Standard Programming Language Fortran 90, ANSI X3.198-1992. v ANSI/IEEE Standard for Binary Floating-Point Arithmetic, ANSI/IEEE Std 754-1985. v Federal (USA) Information Processing Standards Publication Fortran, FIPS PUB 69-1. v Information technology - Programming languages - Fortran, ISO/IEC 1539-1:1991 (E). (This information uses its informal name, Fortran 90.) v Information technology - Programming languages - Fortran - Part 1: Base language, ISO/IEC 1539-1:1997. (This information uses its informal name, Fortran 95.) v Information technology - Programming languages - Fortran - Part 1: Base language, ISO/IEC 1539-1:2004. (This information uses its informal name, Fortran 2003.) v Information technology - Programming languages - Fortran - Part 1: Base language, ISO/IEC 1539-1:2010. (This information uses its informal name, Fortran 2008.)
About this information xiii v Military Standard Fortran DOD Supplement to ANSI X3.9-1978, MIL-STD-1753 (United States of America, Department of Defense standard). Note that XL Fortran supports only those extensions documented in this standard that have also been subsequently incorporated into the Fortran 90 standard. v OpenMP Application Program Interface Version 3.1, available at http://www.openmp.org Other IBM information v Blue Gene/Q Hardware Overview and Installation Planning, SG24-7872, available at http://www.redbooks.ibm.com/redpieces/abstracts/sg247872.html?Open v Blue Gene/Q Hardware Installation and Maintenance Guide, SG24-7974, available at http://www.redbooks.ibm.com/redpieces/abstracts/sg247974.html?Open v Blue Gene/Q High Availability Service Node, REDP-4657, available at http://www.redbooks.ibm.com/redpieces/abstracts/redp4657.html?Open v Blue Gene/Q System Administration, SG24-7869, available at http:// www.redbooks.ibm.com/redpieces/abstracts/sg247869.html?Open v Blue Gene/Q Application Development, SG24-7948, available at http://www.redbooks.ibm.com/redpieces/abstracts/sg247948.html?Open v Blue Gene/Q Code Development and Tools Interface, REDP-4659, available at http://www.redbooks.ibm.com/redpieces/abstracts/redp4659.html?Open
Technical support
Additional technical support is available from the XL Fortran Support page at http://www.ibm.com/software/awdtools/fortran/xlfortran/features/bg/ support/. This page provides a portal with search capabilities to a large selection of Technotes and other support information.
If you cannot find what you need, you can send email to [email protected].
For the latest information about XL Fortran, visit the product information site at http://www.ibm.com/software/awdtools/fortran/xlfortran/features/bg/.
How to send your comments
Your feedback is important in helping to provide accurate and high-quality information. If you have any comments about this information or any other XL Fortran information, send your comments by email to [email protected].
Be sure to include the name of the information, the part number of the information, the version of XL Fortran, and, if applicable, the specific location of the text you are commenting on (for example, a page number or table number).
xiv XL Fortran: Optimization and Programming Guide Chapter 1. Optimizing your applications
The XL compilers enable development of high performance applications by offering a comprehensive set of performance enhancing techniques that exploit the multilayered Blue Gene® architecture. These performance advantages depend on good programming techniques, thorough testing and debugging, followed by optimization, and tuning.
Distinguishing between optimization and tuning You can use optimization and tuning separately or in combination to increase the performance of your application. Understanding the difference between them is the first step in understanding how the different levels, settings, and techniques can increase performance. Optimization
Optimization is a compiler driven process that searches for opportunities to restructure your source code and give your application better overall performance at run time, without significantly impacting development time. The XL compiler optimization suite, which you control using compiler options and directives, performs best on well-written source code that has already been through a thorough debugging and testing process. These optimization transformations can: v Reduce the number of instructions your application executes to perform critical operations. v Restructure your object code to make optimal use of the Blue Gene architecture. v Improve memory subsystem usage. v Exploit the ability of the architecture to handle large amounts of shared memory parallelization.
Each basic optimization technique can result in a performance benefit, although not all optimizations can benefit all applications. Consult the “Steps in the optimization process” on page 2 for an overview of the common sequence of steps you can use to increase the performance of your application. Tuning
While optimization applies general transformations designed to improve the performance of any application in any supported environment, tuning offers you opportunities to adjust specific characteristics or target execution environments of your application to improve its performance. Even at low optimization levels, tuning for your application and target architecture can have a positive impact on performance. With proper tuning the compiler can: v Select more efficient machine instructions. v Generate instruction sequences that are more relevant to your application. v Select from more focussed optimizations to improve your code.
© Copyright IBM Corp. 1990, 2012 1 Steps in the optimization process As you begin the optimization process, consider that not all optimization techniques suit all applications. Trade-offs sometimes occur between an increase in compile time, a reduction in debugging capability, and the improvements that optimization can provide.
Learning about, and experimenting with different optimization techniques can help you strike the right balance for your XL compiler applications while achieving the best possible performance. Also, though it is unnecessary to hand-optimize your code, compiler-friendly programming can be extremely beneficial to the optimization process. Unusual constructs can obscure the characteristics of your application and make performance optimization difficult. Use the steps in this section as a guide for optimizing your application. 1. The Basic optimization step begins your optimization processes at levels 0 and 2. 2. The Advanced optimization step exposes your application to more intense optimizations at levels 3, 4. 3. The High-order transformation (HOT) step can help you limit loop execution time. 4. The Interprocedural analysis (IPA) step can optimize your entire application at once. 5. The Debugging optimized code step can help you identify issues and problems that can occur with optimized code. 6. The Getting more performance section offers other strategies and tuning alternatives to compiler-driven optimization.
The section Compiler-friendly programming techniques contains tips for writing more easily optimized source code.
Basic optimization The XL compiler supports several levels of optimization, with each option level building on the levels below through increasingly aggressive transformations, and consequently using more machine resources.
Ensure that your application compiles and executes properly at low optimization levels before trying more aggressive optimizations. This topic discusses two optimizations levels, listed with complementary options in the Basic optimizations table. The table also includes a column for compiler options that can have a performance benefit at that optimization level for some applications. Table 4. Basic optimizations Additional options Complementary Other options with Optimization level implied by default options possible benefits -O0 -qsimd=auto -qarch -O2 -qmaxmem=8192 -qarch -qmaxmem=-1 -qsimd=auto -qtune -qhot=level=0
Note: Specifying -O without including a level implies -O2.
2 XL Fortran: Optimization and Programming Guide Optimizing at level 0 Benefits at level 0 v Minimal performance improvement, with minimal impact on machine resources. v Exposes some source code problems, helping in the debugging process.
Begin your optimization process at -O0 which the compiler already specifies by default. This level performs basic analytical optimization by removing obviously redundant code, and can result in better compile time. It also ensures your code is algorithmically correct so you can move forward to more complex optimizations. -O0 also includes some redundant instruction elimination and constant folding. The option -qfloat=nofold can be used to suppress folding floating-point operations. Optimizing at this level accurately preserves all debugging information and can expose problems in existing code, such as uninitialized variables.
Additionally, specifying -qarch at this level targets your application for a particular machine and can significantly improve performance by ensuring your application takes advantage of all applicable architectural benefits.
Note: For SMP programs, you need to add an additional option -qsmp=noopt.
For more information on tuning, consult Tuning for Your Target Architecture.
See "-O" in the XL Fortran Compiler Reference for information on the -O level syntax. Optimizing at level 2 Benefits at level 2 v Eliminates redundant code v Basic loop optimization v Can structure code to take advantage of -qarch and -qtune settings
After successfully compiling, executing, and debugging your application using -O0, recompiling at -O2 opens your application to a set of comprehensive low-level transformations that apply to subprogram or compilation unit scopes and can include some inlining. Optimizations at -O2 are a relative balance between increasing performance while limiting the impact on compilation time and system resources. You can increase the memory available to some of the optimizations in the -O2 portfolio by providing a larger value for the -qmaxmem option. Specifying -qmaxmem=-1 allows the optimizer to use memory as needed without checking for limits but does not change the transformations the optimizer applies to your application at -O2. Starting to tune at level 2
Choosing the right hardware architecture target or family of targets becomes even more important at -O2 and higher. By targeting the proper hardware, the optimizer can make the best use of the hardware facilities available. If you choose a family of hardware targets, the -qtune option can direct the compiler to emit code consistent with the architecture choice, but executes optimally on the chosen tuning hardware target. With this option, you can compile for a general set of targets but have the code run best on a particular target.
See the Chapter 2, “Tuning XL compiler applications,” on page 31 topics for details on the -qarch and -qtune options.
Chapter 1. Optimizing your applications 3 The -O2 option can perform a number of additional optimizations, including: v Common subexpression elimination: Eliminates redundant instructions. v Constant propagation: Evaluates constant expressions at compile-time. v Dead code elimination: Eliminates instructions that a particular control flow does not reach, or that generate an unused result. v Dead store elimination: Eliminates unnecessary variable assignments. v Graph coloring register allocation: Globally assigns user variables to registers. v Value numbering: Simplifies algebraic expressions, by eliminating redundant computations. v Instruction scheduling for the target machine. v Loop unrolling and software pipelining. v Moving invariant code out of loops. v Simplifying control flow. v Strength reduction and effective use of addressing modes.
Even with -O2 optimizations, some useful information about your source code is made available to the debugger if you specify -g. Using a higher -g level increases the information provided to the debugger, but reduces the optimization that can be done. Conversely, higher optimization levels can transform code to an extent to which debugging information is no longer accurate. Use that information with discretion.
The section on “Debugging optimized code” on page 20 discusses other debugging strategies in detail.
See "-O" in the XL Fortran Compiler Reference for information on the -O level syntax.
Advanced optimization Higher optimization levels can have a tremendous impact on performance, but some trade-offs can occur in terms of code size, compile time, resource requirements, and numeric or algorithmic precision.
After applying “Basic optimization” on page 2 and successfully compiling and executing your application, you can apply more powerful optimization tools. The XL compiler optimization portfolio includes many options for directing advanced optimization, and the transformations your application undergoes are largely under your control. The discussion of each optimization level in Table 5 includes information on not only the performance benefits, and the possible trade-offs as well, but information on how you can help guide the optimizer to find the best solutions for your application. Table 5. Advanced optimizations Additional options Complementary Options with Optimization Level implied options possible benefits -O3 -qnostrict -qarch -qmaxmem=-1 -qtune -qhot=level=0 -qsimd=auto
4 XL Fortran: Optimization and Programming Guide Table 5. Advanced optimizations (continued) Additional options Complementary Options with Optimization Level implied options possible benefits -O4 -qnostrict -qarch -qsmp=auto -qmaxmem=-1 -qtune -qhot -qcache -qipa -qarch=auto -qtune=auto -qcache=auto -qsimd=auto -O5 All of -O4 -qarch -qsmp=auto -qipa=level=2 -qtune -qcache
When you compile programs with any of the following sets of options: v -qhot -qnostrict v -qhot -O3 v -O4 v -O5 the compiler automatically attempts to vectorize calls to system math functions by calling the equivalent vector functions in the Mathematical Acceleration Subsystem libraries (MASS), with the exceptions of functions vatan2, vsatan2, vdnint, vdint, vcosisin, vscosisin, vqdrt, vsqdrt, vrqdrt, vsrqdrt, vpopcnt4, and vpopcnt8.Ifthe compiler cannot vectorize, it automatically tries to call the equivalent MASS scalar functions. For automatic vectorization or scalarization, the compiler uses versions of the MASS functions contained in the system library libxlopt.a.
In addition to any of the preceding sets of options, when the -qipa option is in effect, if the compiler cannot vectorize, it tries to inline the MASS scalar functions before deciding to call them. Optimizing at level 3 Benefits at level 3 v Automatic generation of SIMD instructions (-qhot=level=0) v In-depth “Aliasing” on page 41 analysis v Better loop scheduling v High-order loop analysis and transformations (-qhot=level=0) v Inlining of small procedures within a compilation unit by default v Eliminating implicit compile-time memory usage limits v Widening, which merges adjacent load/stores and other operations v Pointer aliasing improvements to enhance other optimizations
Specifying -O3 initiates more intense low-level transformations that remove many of the limitations present at -O2. For instance, the optimizer no longer checks for memory limits, by defaulting to -qmaxmem=-1. Additionally, optimizations encompass larger program regions and attempt more in-depth analysis. While not all applications contain opportunities for the optimizer to provide a measurable increase in performance, most applications can benefit from this type of analysis.
Chapter 1. Optimizing your applications 5 Potential trade-offs at level 3
With the in-depth analysis of -O3 comes a trade-off in terms of compilation time and memory resources. Also, since -O3 implies -qnostrict, the optimizer can alter certain floating-point semantics in your application to gain execution speed. This typically involves precision trade-offs as follows: v Reordering of floating-point computations. v Reordering or elimination of possible exceptions, such as division by zero or overflow. v Using alternative calculations that might give slightly less precise results or not handle infinities or NaNs in the same way.
You can still gain most of the -O3 benefits while preserving precise floating-point semantics by specifying -qstrict. Compiling with -qstrict is necessary if you require the same absolute precision in floating-point computational accuracy as you get with -O0, -O2,or-qnoopt results. The option -qstrict=ieeefp also ensures adherence to all IEEE semantics for floating-point operations. If your application is sensitive to floating-point exceptions or the order of evaluation for floating-point arithmetic, compiling with -qstrict, -qstrict=exceptions,or-qstrict=order helps to ensure accurate results. You should also consider the impact of the -qstrict=precision suboption group on floating-point computational accuracy. The precision suboption group includes the individual suboptions: subnormals, operationprecision, association, reductionorder, and library (described in the -qstrict option in the XL Fortran Compiler Reference).
Without -qstrict, the difference in computation for any one source-level operation is very small in comparison to “Basic optimization” on page 2. Although a small difference can be compounded if the operation is in a loop structure where the difference becomes additive, most applications are not sensitive to the changes that can occur in floating-point semantics.
See "-O" in the XL Fortran Compiler Reference for information on the -O level syntax. An intermediate step: adding -qhot suboptions at level 3 At -O3, the optimization includes minimal -qhot loop transformations at level=0 to increase performance. You can further increase your performance benefit by increasing the level and therefore the aggressiveness of -qhot. Try specifying -qhot without any suboptions, or -qhot=level=1.
The following -qhot suboptions can also provide additional performance benefits, depending on the characteristics of your application: v -qhot=vector to enable long vectorization v -qhot=arraypad to enable array padding v -qhot=fastmath to enable the replacement of math routines with those from the XLOPT library
For more information on -qhot, see “High-order transformation (HOT)” on page 9.
Conversely, if the application does not use loops processing arrays (which -qhot improves), you can improve compile speed with minimal performance loss by using -qnohot after -O3.
6 XL Fortran: Optimization and Programming Guide Optimizing at level 4 Benefits at level 4 v Propagation of global and argument values between compilation units v Inlining code from one compilation unit to another v Reorganization or elimination of global data structures v An increase in the precision of aliasing analysis
Optimizing at -O4 builds on -O3 by triggering -qipa=level=1 which performs interprocedural analysis (IPA), optimizing your entire application as a unit. This option is particularly pertinent to applications that contain a large number of frequently used routines.
To make full use of IPA optimizations, you must specify -O4 on the compilation and link steps of your application build as interprocedural analysis occurs in stages at both compile and link time. Potential trade-offs at level 4
In addition to the trade-offs already mentioned for -O3, specifying -qipa can significantly increase compilation time, especially at the link step.
See "-O" in the XL Fortran Compiler Reference for information on the -O level syntax. The IPA process 1. At compile time optimizations occur on a file-by-file basis, as well as preparation for the link stage. IPA writes analysis information directly into the object files the compiler produces. 2. At the link stage, IPA reads the information from the object files and analyzes the entire application. 3. This analysis guides the optimizer on how to rewrite and restructure your application and apply appropriate -O3 level optimizations.
The “Interprocedural analysis (IPA)” on page 11 section contains more information on IPA including details on IPA suboptions.
Beyond -qipa, -O4 enables other optimization options: v -qhot Enables more aggressive HOT transformations to optimize loop constructs and array language. v -qhot=vector Optimizes array data to run mathematical operations in parallel where applicable. v -qarch=auto and -qtune=auto Optimizes your application to execute on a hardware architecture identical to your build machine. If the architecture of your build machine is incompatible with your application's execution environment, you must specify a different -qarch suboption after the -O4 option. This overrides -qarch=auto. v -qcache=auto Optimizes your cache configuration for execution on specific hardware architecture. The auto suboption assumes that the cache configuration of your build machine is identical to the configuration of your execution architecture.
Chapter 1. Optimizing your applications 7 Specifying a cache configuration can increase program performance, particularly loop operations by blocking them to process only the amount of data that can fit into the data cache. If you want to execute your application on a different machine, specify correct cache values. Optimizing at level 5 Benefits at level 5 v Most aggressive optimizations available v Makes full use of loop optimizations and “Interprocedural analysis (IPA)” on page 11
As the highest optimization level, -O5 includes all -O4 optimizations and deepens whole program analysis by increasing the -qipa level to 2. Compiling with -O5 also increases how aggressively the optimizer pursues aliasing improvements. Additionally, if your application contains a mix of C/C++ and Fortran code that you compile using the XL compilers, you can increase performance by compiling and linking your code with the -O5 option. Potential trade-offs at level 5
Compiling at -O5 requires more compile time and machine resources than any other optimization levels, particularly if you include -O5 on the IPA link step. Compile at -O5 as the final phase in your optimization process after successfully compiling and executing your application at -O4.
See "-O" in the XL Fortran Compiler Reference for information on the -O level syntax.
Specialized optimization techniques While some optimization techniques are active at advanced optimization levels, certain types of applications can receive a performance benefit even when you apply only basic optimizations. Table 6. Specialized optimization techniques Technique Benefit HOT Minimizes loop execution time which is beneficial to most applications that contain large loops, or many small loops. HOT also improves memory access patterns in your application. IPA Performs whole program analysis, providing the optimization suite with a complete view of your entire application. This applies performance enhancements with more focus and robustness. Vector technology Vector technology is a PowerPC® technology for accelerating the performance-driven, high-bandwidth communications and computing applications. You can use the vector technology to get dramatic performance improvement for your applications.
8 XL Fortran: Optimization and Programming Guide Table 6. Specialized optimization techniques (continued) Technique Benefit Compiler reports You can use the -qlistfmt option to generate a compiler report in XML 1.0 format that indicates some of the details of how your program was optimized. You can use this information to understand your application code and to tune your code for better performance.
High-order transformation (HOT) As part of the XL compiler optimization suite, the HOT transformations focus specifically on loops which typically account for the majority of the execution time for most applications. HOT transformations perform in-depth loop analysis to minimize their execution time.
Loop optimization analysis includes: v Interchange v Fusion v Unrolling loop nests v Reducing the use of temporary arrays
The goals of these optimizations include: v Reducing memory access costs through effective cache use and translation look-aside buffers (TLBs). Increasing memory locality reduces cache and TLB misses. v Overlapping computation and memory access through effective utilization of the hardware data prefetching capabilities. v Improving processor resource utilization by reordering and balancing the use of instructions with complementary resource requirements. Loop computation balance typically involves creating an equitable relationship between load/store operations and floating-point computations.
Compiling with -O3 and higher triggers HOT transformations by default. You can also see performance benefits by specifying -qhot with -O2, or adding more -qhot optimizations than the default level=0 at -O3 .
You can see particular -qhot benefits if your application contains Fortran 90-style array language constructs, as HOT transformations include elimination of intermediate temporary variables and statement fusion.
You can also use directives to assist in loop analysis. Assertive directives such as INDEPENDENT or CNCALL allow you to describe important loop characteristics or behaviors that HOT transformations can exploit. Prescriptive directives such as UNROLL or PREFETCH allow you to direct the HOT transformations on a loop-by-loop basis. You can also specify the -qreport compiler option to generate information about loop transformations. The report can assist you in deciding where best to include directives to improve the performance of your application. For example, you can use this section of the listing to identify non-stride-one references that may prevent loop vectorization.
You can use the -qreport option in conjunction with -qhot or any optimization option that works with -qhot to produce a pseudo-Fortran report showing how the
Chapter 1. Optimizing your applications 9 loops were transformed. The LOOP TRANSFORMATION SECTION of the listing file also contains information about data prefetch insertion locations.
When used with -qsmp, -qhot=level=2 instructs the compiler to perform the transformations of -qhot=level=1 plus some additional transformation on nested loops. The resulting loop analysis and transformations can lead to more cache reuse and loop parallelization. If you use -qhot=level=2 and -qsmp together with -qreport or -qlistfmt, you can see this information on aggressive loop analysis performed on loop nests in the LOOP TRANSFORMATION SECTION of the listing file or compiler report.
With the -qassert=refalign suboption, the compiler might generate more efficient code. This assertion is particularly useful when you target a Single Instruction Multiple Data (SIMD) architecture with -qhot=level=0 or -qhot=level=1 with the -qsimd=auto option.
In addition to general loop transformation, -qhot supports suboptions that you can specify to enable additional transformations detailed in this section. HOT short vectorization
When you are targeting a PowerPC processor that supports Quad Processing Extension (QPX) vectors, the default -qsimd=auto option enables the compiler to transform code into QPX vector instructions. These machine instructions can execute up to sixteen operations in parallel. This transformation mostly applies to the loops that iterate over contiguous array data and perform calculations on each element. You can use the NOSIMD directive to prevent the transformation of a particular loop. HOT long vectorization
When you specify any of the following: v -O4 and higher v -qhot with -qnostrict
you enable -qhot=vector by default. Specifying -qnostrict with optimizations other than -O4 and -O5 ensures that the compiler looks for long vectorization opportunities. This can optimize loops in source code for operations on array data by ensuring that operations run in parallel where applicable. The compiler uses standard machine registers for these transformations and does not restrict vector data size; supporting both single- and double-precision floating-point vectorization. Often, HOT vectorization involves transformations of loop calculations into calls to specialized mathematical routines supplied with the compiler such as the Mathematical Acceleration Subsystem (MASS) libraries. These mathematical routines use algorithms that calculate results more efficiently than executing the original loop code.
For more information about optimization levels like -O4 and the other compiler options they imply, see “Advanced optimization” on page 4. HOT array size adjustment
An array dimension that is a power of two can lead to a decrease in cache utilization. The -qhot=arraypad suboption allows the compiler to increase the dimensions of arrays where doing so could improve the efficiency of array-processing loops. Using this suboption can reduce cache misses and page
10 XL Fortran: Optimization and Programming Guide faults that slow your array processing programs. The HOT transformations will not necessarily pad all arrays, and can pad different arrays by different amounts in order to gain performance. You can specify a padding factor to apply to all arrays. This value is typically a multiple of the largest array element size.
Use -qhot=arraypad with discretion as array padding uses more memory and the performance trade-off does not benefit all applications. Also, these HOT transformations do not include checks for array data overlay, as with Fortran EQUIVALENCE, or array shaping operations. HOT fast scalar math routines
The XLOPT library contains faster versions of certain math functions that are normally provided by the operating system or in the default runtime. With -qhot=fastmath, the compiler replaces calls to the math functions with their faster counterparts in XLOPT library. This option requires -qstrict=nolibrary in effect. Interprocedural analysis (IPA) Interprocedural Analysis (IPA) can analyze and optimize your application as a whole, rather than on a file-by-file basis.
Run during the link step of an application build, the entire application, including linked libraries, is available for interprocedural analysis. This whole program analysis opens your application to a powerful set of transformations available only when more than one file or compilation unit is accessible. IPA optimizations are also effective on mixed language applications.
PDF info Libraries
IPA Objects IPA
System EXE Other Objects Partitions Linker DLL
Low-level optimizer
Optimized Objects
Figure 1. IPA at the link step
The following are some of the link-time transformations that IPA can use to restructure and optimize your application: v Inlining between compilation units. v Complex data flow analyses across subprogram calls to eliminate parameters or propagate constants directly into called subprograms.
Chapter 1. Optimizing your applications 11 v Improving parameter usage analysis, or replacing external subprogram calls to system libraries with more efficient inline code. v Restructuring data structures to maximize access locality.
In order to maximize IPA link-time optimization, you must use IPA at both the compile and link step. Objects you do not compile with IPA can only provide minimal information to the optimizer, and receive minimal benefit. However when IPA is active on the compile step, the resulting object file contains program information that IPA can read during the link step. The program information is invisible to the system linker, and you can still use the object file and link without invoking IPA. The IPA optimizations use hidden information to reconstruct the original compilation and can completely analyze the subprograms the object contains in the context of their actual usage in your application.
During the link step, IPA restructures your application, partitioning it into distinct logical code units. After IPA optimizations are complete, IPA applies the same low-level compilation-unit transformations as the -O2 and -O3 base optimizations levels. Following those transformations, the compiler creates one or more object files and linking occurs with the necessary libraries through the system linker.
It is important that you specify a set of compilation options as consistent as possible when compiling and linking your application. This includes all compiler options, not just -qipa suboptions. When possible, specify identical options on all compilations and repeat the same options on the IPA link step. Incompatible or conflicting options that you specify to create object files, or link-time options in conflict with compile-time options can reduce the effectiveness of IPA optimizations. Using IPA on the compile step only About this task
IPA can still perform transformations if you do not specify IPA on the link step. Using IPA on the compile step initiates optimizations that can improve performance for an individual object file even if you do not link the object file using IPA. The primary focus of IPA is link-step optimization, but using IPA only on the compile-step can still be beneficial to your application without incurring the costs of link-time IPA.
12 XL Fortran: Optimization and Programming Guide C Front End C++ Front End Fortran Front End
Array Language Processor
IPA
Low-level optimizer
IPA Object
Figure 2. IPA at the compile step
IPA Levels and other IPA suboptions
You can control many IPA optimization functions using the -qipa option and suboptions. The most important part of the IPA optimization process is the level at which IPA optimization occurs. Default compilation does not invoke IPA. If you specify -qipa without a level, or specify -O4, IPA optimizations are at level one. If you specify -O5, IPA optimizations are at level two. Table 7. The levels of IPA IPA Level Behaviors qipa=level=0 v Automatically recognizes standard library functions v Localizes statically bound variables and procedures v Organizes and partitions your code according to call affinity, expanding the scope of the -O2 and -O3 low-level compilation unit optimizer v Lowers compilation time in comparison to higher levels, though limits analysis qipa=level=1 v Level 0 optimizations v Performs procedure inlining across compilation units v Organizes and partitions static data according to reference affinity
Chapter 1. Optimizing your applications 13 Table 7. The levels of IPA (continued) IPA Level Behaviors
qipa=level=2 v Level 0 and level 1 optimizations v Performs whole program alias analysis which removes ambiguity between pointer references and calls, while refining call side effect information v Propagates interprocedural constants v Eliminates dead code v Performs pointer analysis v Performs procedure cloning v Optimizes intraprocedural operations, using specifically: – Value numbering – Code propagation and simplification – Code motion, into conditions and out of loops – Redundancy elimination techniques v Performs data reorganization
IPA includes many suboptions that can help you guide IPA to perform optimizations important to the particular characteristics of your application. Among the most relevant to providing information on your application are: v lowfreq, with which you can specify a list of procedures that are likely to be called infrequently during the course of a typical program run. Performance can increase because optimization transformations will not focus on these procedures. v partition, with which you can specify the size of the regions within the program to analyze. Larger partitions contain more procedures, which result in better interprocedural analysis but require more storage to optimize. v threads, with which you can specify the number of parallel threads available to IPA optimizations. This can provide an increase in compilation-time performance on multi-processor systems. Using IPA across the XL compiler family About this task
The XL compiler family shares optimization technology. Object files you create using IPA on the compile step with the XL C, C++, and Fortran compilers can undergo IPA analysis during the link step. Where program analysis shows that objects were built with compatible options, such as -qnostrict, IPA can perform transformations such as inlining C functions into Fortran code, or propagating C++ constant data into C function calls. Vector technology Vector technology is a PowerPC technology for accelerating the performance-driven, high-bandwidth communications and computing applications. You can use the vector technology to get dramatic performance improvement for your applications.
There are two ways of using the vector technology: hand coding and automatic vectorization. Automatic vectorization often brings the best performance when you write the code in the right way, but appropriate hand coding can provide additional performance improvement.
14 XL Fortran: Optimization and Programming Guide The following example shows the difference between a simple array element addition and a vectorized version of the same loop.
Array element addition without using the vector technology: subroutine myadd(n) integer :: i, n real(8), dimension(n) :: a, b, c
do i=1, n a(i) = b(i) + c(i) enddo end subroutine
Modified array element addition utilizing the vector technology: subroutine myadd_vector(n) integer :: j, n ! vector_size is a constant vector(real(8)), dimension(n/vector_size) :: v_a, v_b, v_c
do j=1, n/vector_size v_a(j) = vec_add(v_b(j), v_c(j)) enddo end subroutine
In the vectorized version of the code, the data type is replaced by the vector data type. The loop range is reduced from n to n/vector_size. Without the vector technology, multiple instructions cost many processor clock cycles. With the vector technology, the operation, v_a(j)=vec_add(v_b(j), v_c(j)), is executed in a single machine instruction for each vector. Therefore, the vector technology can improve the performance of an application.
This section provides general information about vector technology with the following three subsections: v “Vector technology information” v “Explicitly calling vector libraries for vectorization” on page 17 v “Auto-vectorization limitations” on page 17 Vector technology information
This section provides links to all of the information about the vector technology and categorize them into the following types: v Using vector technology with hand coding v Using vector technology with auto-vectorization
Using vector technology with hand coding
The following table lists the information about using the vector technology with hand coding and provides the links to the detailed information in different documents. Table 8. Language features for using vector technology with hand coding: Information you need Sections you can read Intrinsic data types Vector (IBM extension) in XL Fortran Language Reference Vector type declaration statement Vector (IBM extension) in XL Fortran Language Reference
Chapter 1. Optimizing your applications 15 Table 8. Language features for using vector technology with hand coding: (continued) Information you need Sections you can read Vector intrinsic procedures Vector intrinsic procedures (IBM extension) in XL Fortran Language Reference Using the vector libraries Using the vector libraries
Using vector technology with auto-vectorization
The following table lists the information about compiler options for auto-vectorization and provides the links to the detailed information in different documents. Table 9. Information about compiler options for auto-vectorization To do... Read... Enable generation of vector instructions for -qsimd in XL Fortran Compiler Reference processors that support them. Perform high-order transformations (HOT) -qhot in XL Fortran Compiler Reference during optimization.
Produce listing files and understand how v -qlistfmt in XL Fortran Compiler Reference sections of code have been optimized. v -qreport in XL Fortran Compiler Reference v Using compiler reports to diagnose optimization opportunities v Parsing compiler reports with development tools Ensure that optimizations done by default, -qstrict in XL Fortran Compiler Reference do not alter certain program semantics related to strict IEEE floating-point conformance.
Tuning for your target architecture using v Tuning for your target architecture -qarch and -qtune. v Using -qtune
The following table lists the directive and compiler option that you can use to prohibit auto-vectorization and provides the links to the detailed information in different documents. Table 10. Directive and compiler option for auto-vectorization To do... Read... Prohibit the compiler from auto-vectorizing NOVECTOR in XL Fortran Language the loop immediately following the directive. Reference Disable auto-vectorization. -qsimd in XL Fortran Compiler Reference
Some optimization processes are related to auto-vectorization, you can use compiler options to control these optimizations. The following table lists these optimization processes and provides the links to the detailed information in different documents.
16 XL Fortran: Optimization and Programming Guide Table 11. Optimizations related to auto-vectorization To learn about... Read...
The High-order transformation (HOT) v High-order transformation (HOT) v An intermediate step: adding -qhot suboptions at level 3 The Interprocedural analysis (IPA) The IPA process
Explicitly calling vector libraries for vectorization
To use the vector technology in your applications, you can either rewrite the algorithm manually or rely on the automatic vectorization of the compiler. Although automatic vectorization can provide the highest performing solution, proper hand coding can also bring good performance.
The following example shows how to explicitly call the vector libraries to make use of the vector functionality provided by the target hardware. function dotp(x,y,n) result(s) real*8 x(*),y(*),s vector(real(8)) sv,xv,yv integer i,n
sv = vec_splats(0.0D0) do i=1,n,2 xv = vec_ld2(0,x(i)) yv = vec_ld2(0,y(i)) sv = vec_madd(xv,yv,sv) enddo s = vec_extract(sv,0)+vec_extract(sv,1) if (mod(n,2) .eq. 1) then s=s+x(n)*y(n) endif end function program dot real*8 x(100),y(100),s integer i do i=1,100 x(i)=0.5*i y(i)=2.0 enddo s = dotp(x,y,100) print *,s end
The program performs the dot product for two arrays of REAL. At each iteration, two elements from the arrays are loaded into two REAL vector variables. The program then uses a multiply add operation to calculate the product of the two vectors and add the product with the previous sum. At the end of the loop the two elements of the vector that hold the partial sums are added to form the complete sum value. If the size of the input vectors do not evenly fit in the vector variables, a single scalar product is performed to complete the dot product computation. Auto-vectorization limitations
When you use the auto-vectorization, you might find that some transformations cannot be performed. If you compile with -qhot and -qlistfmt=xml=transforms or -qlistfmt=xml=all, you can get a compiler report that lists the reasons why some
Chapter 1. Optimizing your applications 17 transformations were not performed. For detailed information about the possible reasons, see Using compiler reports to diagnose optimization opportunities.
This section uses two code examples to illustrate why auto-vectorization cannot be performed under certain situations.
Example 1: program try real*8 x(100) integer i x(1)=9 do i=2,100 x(i)=x(i-1) enddo end
The x(i)=x(i-1) statement violates the restriction that "a loop cannot be automatically parallelized if one of its variable carries a dependency". x(i) or x(i-1) depends on each other in this sample, which makes the loop non-vectorizable.
Example 2: program try real*8 x(100) integer i do i=1,100,5 x(i)=i + 8; x(i+1)=i + 9; x(i+2)=i + 12; x(i+3)=i + 15; enddo end
The following statements violate the restriction that auto-vectorization cannot be performed if the loop contains a non stride one store. x(i)=i + 8; x(i+1)=i + 9; x(i+2)=i + 12; x(i+3)=i + 15;
In each iteration of the loop, four elements in the array x are accessed and one element is skipped. This continues until the end of the loop, which makes the loop cannot be vectorized. Using compiler reports to diagnose optimization opportunities You can use the -qlistfmt option to generate a compiler report in XML or HTML format that indicates some of the details of how your program was optimized. You can also use the genhtml tool to convert an existing XML report to HTML format. This information can be used to understand your application code and to tune your code for better performance.
The compiler report in XML format can be viewed in a browser that supports XSLT. If you compile with the stylesheet suboption, -qlistfmt=xml=all:stylesheet=xlstyle.xsl, the report contains a link to a stylesheet that renders the XML readable and provides you with opportunities to improve the optimization of your code. You can also create tools to parse this information.
18 XL Fortran: Optimization and Programming Guide Inline reports
If compiled with -qinline and one of -qlistfmt=xml=inlines, -qlistfmt=html=inlines, -qlistfmt=xml or -qlistfmt=html, the compiler report that is generated includes a list of inline attempts during the compilation. The report also specifies the type of attempt and its outcome.
For each function that the compiler has attempted to inline, there is an indication of whether the inline was successful. The report might contain any number of explanations for a named function that has not been successfully inlined. Some examples of these explanations are: v FunctionTooBig - The function is too big to be inlined. v RecursiveCall - The function is not inlined because it is recursive. v ProhibitedByUser - Inlining was not performed because of a user specified pragma or directive. v CallerIsNoopt - No inlining was performed because the caller was compiled without optimization. v WeakAndNotExplicitlyInline - The calling function is weak and not marked as inline.
For a complete list of the possible explanations, see the Inline optimization types section of the XML schema help file called XMLContent.html in the /opt/ibmcmp/xlf/bg/14.1/listings/ directory, which also includes its Japanese and Chinese version, XMLContent-Japanese.utf8.html and XMLContent- Chinese.utf8.html. Loop transformations
If compiled with -qhot and one of -qlistfmt=xml=transforms, -qlistfmt=html=transforms, -qlistfmt=xml or -qlistfmt=html, the compiler report that is generated includes a list of the transformations performed on all loops in the file during the compilation. It also lists reasons why some transformations were not performed. v Reasons why a loop cannot be automatically parallelized v Reasons why a loop cannot be unrolled v Reasons why SIMD vectorization failed
For a complete list of the possible transformation problems, see the Loop transformation types section of the XML schema help file called XMLContent.html in the /opt/ibmcmp/xlf/bg/14.1/listings/ directory, which also includes its Japanese and Chinese version, XMLContent-Japanese.utf8.html and XMLContent-Chinese.utf8.html. Data reorganizations
If compiled with -qhot and one of -qlistfmt=xml=data, -qlistfmt=html=data, -qlistfmt=xml or -qlistfmt=html, the compiler report that is generated includes a list of data reorganizations performed on the program during compilation. Here are some examples of data reorganizations: v Array splitting v Array coalescing v Array interleaving v Array transposition
Chapter 1. Optimizing your applications 19 v Common block splitting v Memory merge
For each of these reorganizations, the report contains details about the name of the data, file names, line numbers, and the region names. Parsing compiler reports with development tools Software development tools can be created to parse the compiler reports produced in XML format. These tools can help direct you to opportunities to improve the performance of your application.
The compiler includes an XML schema that you can use to create a tool to parse the compiler reports and display aspects of your code that may represent performance improvement opportunities. The schema, xllisting.xsd, is located in the /opt/ibmcmp/xlf/bg/14.1/listings/ directory. This schema helps to present the information from the report in a tree structure.
You can also find a schema help file called XMLContent.html that helps you understand the details of the schema.
Debugging optimized code Debugging optimized programs presents special usability problems. Optimization can change the sequence of operations, add or remove code, change variable data locations, and perform other transformations that make it difficult to associate the generated code with the original source statements.
For example: Data location issues With an optimized program, it is not always certain where the most current value for a variable is located. For example, a value in memory may not be current if the most current value is being stored in a register. Most debuggers are incapable of following the removal of stores to a variable, and to the debugger it appears as though that variable is never updated, or possibly even set. This contrasts with no optimization where all values are flushed back to memory and debugging can be more effective and usable. Instruction scheduling issues With an optimized program, the compiler may reorder instructions. That is, instructions may not be executed in the order the programmer would expect based on the sequence of lines in their original source code. Also, the sequence of instructions may not be contiguous. As the user steps through their program with a debugger, it may appear as if they are returning to a previously executed line in their code (interleaving of instructions). Consolidating variable values Optimizations can result in the removal and consolidation of variables. For example, if a program has two expressions that assign the same value to two different variables, the compiler may substitute a single variable. This can inhibit debug usability because a variable that a programmer is expecting to see is no longer available in the optimized program.
There are a couple of different approaches you can take to improve debug capabilities while also optimizing your program:
20 XL Fortran: Optimization and Programming Guide Debug non-optimized code first Debug a non-optimized version of your program first, then recompile it with your desired optimization options. See “Debugging in the presence of optimization” for some compiler options that are useful in this approach. Use -g level Use the -g level suboption to control the amount of debugging information made available. Increasing it improves debug capability, but prevents some optimizations. Use -qoptdebug When compiling with -O3 optimization or higher, use the compiler option -qoptdebug to generate a pseudocode file that more accurately maps to how instructions and variable values will operate in an optimized program. With this option, when you load your program into a debugger, you will be debugging the pseudocode for the optimized program. See “Using -qoptdebug to help debug optimized programs” on page 22 for more information. Understanding different results in optimized programs Here are some reasons why an optimized program might produce different results from one that has not undergone the optimization process: v Optimized code can fail if a program contains code that is not valid. For example, failure can occur if the program passes an actual argument that also appears in a common block in the called procedure, or if two or more dummy arguments are associated with the same actual argument. The optimization process relies on your application conforming to language standards. v If a program that works without optimization fails when you optimize, check the cross-reference listing and the execution flow of the program for variables that are used before they are initialized. Compile with the -qinitauto=hex_value or -qinitalloc=hex_value option to try to produce the incorrect results consistently. For example, using -qinitauto=FF gives REAL and COMPLEX variables an initial value of "negative not a number" (-NAN). Any operations on these variables will also result in NAN values. Other bit patterns (hex_value) may yield different results and provide further clues as to what is going on. Programs with uninitialized variables can appear to work properly when compiled without optimization, because of the default assumptions the compiler makes, but can fail when you optimize. Similarly, a program can appear to execute correctly after optimization, but fails at lower optimization levels or when run in a different environment. v A variation on uninitialized storage. Referring to an automatic-storage variable by its address after the owning function has gone out of scope leads to a reference to a memory location that can be overwritten as other auto variables come into scope as new functions are called.
Use with caution debugging techniques that rely on examining values in storage. The compiler might have deleted or moved a common expression evaluation. It might have assigned some variables to registers, so that they do not appear in storage at all. Debugging in the presence of optimization Debug and compile your program with your desired optimization options. Test the optimized program before placing it into production. If the optimized code does not produce the expected results, you can attempt to isolate the specific optimization problems in a debugging session.
Chapter 1. Optimizing your applications 21 The following list presents options that provide specialized information, which can be helpful during the development of optimized code: -qlist Instructs the compiler to emit an object listing. The object listing includes hex and pseudo-assembly representations of the generated instructions, traceback tables, and text constants. -qreport Instructs the compiler to produce a report of the loop transformations it performed and how the program was parallelized. For -qreport to generate a listing, the options -qhot or -qsmp should also be specified. -qipa=list Instructs the compiler to emit an object listing that provides information for IPA optimization. -qcheck Generates code that performs certain types of runtime checking. -qsmp=noopt If you are debugging SMP code, -qsmp=noopt ensures that the compiler performs only the minimum transformations necessary to parallelize your code and preserves maximum debug capability. -qoptdebug When used with high levels of optimization, produces files containing optimized pseudocode that can be read by a debugger. -qkeepparm Ensures that procedure parameters are stored on the stack even during optimization. This can negatively impact execution performance. The -qkeepparm option then provides access to the values of incoming parameters to tools, such as debuggers, simply by preserving those values on the stack. -qinitalloc Instructs the compiler to emit code that initializes all allocatable and pointer variables that are allocated but not initialized to a given value. -qinitauto Instructs the compiler to emit code that initializes all automatic variables to a given value. -g, -qdbg Generates debugging information for use by a symbolic debugger. You can use different -g or -qdbg levels to debug optimized code by viewing or possibly modifying accessible variables at selected source locations in the debugger. Higher -g or -qdbg levels provide a more complete debug support, while lower levels provide higher runtime performance. For details, see -g or -qdbg.
In addition, you can also use the SNAPSHOT directive to ensure that certain variables are visible to the debugger at points in your application. For details, see SNAPSHOT. Using -qoptdebug to help debug optimized programs The purpose of the -qoptdebug compiler option is to aid the debugging of optimized programs. It does this by creating pseudocode that maps more closely to the instructions and values of an optimized program than the original source code. When a program compiled with this option is loaded into a debugger, you will be
22 XL Fortran: Optimization and Programming Guide debugging the pseudocode rather than your original source. By making optimizations explicit in pseudocode, you can gain a better understanding of how your program is really behaving under optimization. Files containing the pseudocode for your program are generated with the file suffix .optdbg. Only line debugging is supported for this feature.
Compile your program as in the following example: bgxlf95 myprogram.f -O3 -qhot -g -qoptdebug -o myprogram
In this example, your source file is compiled to a.out. The pseudocode for the optimized program is written to a file called myprogram.optdbg which can be referred to while debugging your program.
Notes: v The -g or the -qlinedebug option must also be specified in order for the compiled executable to be debuggable. However, if neither of these options are specified, the pseudocode file
From the following examples, you can see how the compiler might apply optimizations to a simple program and how debugging it can differ from debugging your original source.
Example 1: Represents the original non-optimized code for a simple program. It presents an optimization opportunity to the compiler; the loop can be unrolled. In the optimized source, you can see iterations of the loop listed explicitly.
Example 2: Represents a listing of the optimized source as shown in the debugger. You can see the unrolled loop in the optimized source.
Example 3: Shows an example of stepping through the optimized source using the debugger. Note, there is no longer a correspondence between the line numbers for these statements in the optimized source as compared to the line numbers in the original source.
Example 1: Original code FUNCTION FOO(X, Y) RESULT(R) IMPLICIT NONE REAL :: X, Y, Z, R INTEGER :: I Z=X+Y DOI=1,4 PRINT *, Z + REAL(I) END DO R=Z END FUNCTION
PROGRAM MAIN IMPPLICIT NONE
Chapter 1. Optimizing your applications 23 REAL :: X, FOO X = FOO(3.0, 4.0) PRINT *, X END PROGRAM MAIN
Example 2: gdb debugger listing (gdb) list 1,55 1 2 3 1| REAL*4 FUNCTION foo (x, y) 4 8| #2 = _xlfBeginIO(",257,#1,32768,NULL,0,NULL) 5 T_2 = (x + y) + 1.00000000E+00 6 CALL _xlfWriteLDReal(%VAL(#2),T_2,",") 7 _xlfEndIO(%VAL(#2)) 8 #2.rnn5 = _xlfBeginIO(",257,#1,32768,NULL,0,NULL) 9 T_2 = (x + y) + 2.00000000E+00 10 CALL _xlfWriteLDReal(%VAL(#2.rnn5),T_2,",") 11 _xlfEndIO(%VAL(#2.rnn5)) 12 #2.rnn4 = _xlfBeginIO(",257,#1,32768,NULL,0,NULL) 13 T_2 = (x + y) + 3.00000000E+00 14 CALL _xlfWriteLDReal(%VAL(#2.rnn4),T_2,",") 15 _xlfEndIO(%VAL(#2.rnn4)) 16 #2.rnn3 = _xlfBeginIO(",257,#1,32768,NULL,0,NULL) 17 T_2 = (x + y) + 4.00000000E+00 18 CALL _xlfWriteLDReal(%VAL(#2.rnn3),T_2,",") 19 _xlfEndIO(%VAL(#2.rnn3)) 20 11| RETURN 21 END FUNCTION foo 22 23 24 13| PROGRAM main () 25 17| T_4 = 3.00000000E+00 26 T_5 = 4.00000000E+00 27 x = foo(T_4,T_5) 28 19| #2 = _xlfBeginIO(",257,#1,32768,NULL,0,NULL) 29 #3 = x 30 CALL _xlfWriteLDReal(%VAL(#2),#3,",") 31 _xlfEndIO(%VAL(#2)) 32 20| CALL _xlfExit(0) 33 END PROGRAM main
Example 3: Stepping through optimized source (gdb) break foo Breakpoint 1 at 0x100006a0: file myprogram.o.optdbg, line 3. (gdb) run Starting program: /home/user/myprogram [Thread debugging using libthread_db enabled] [New Thread -134438912 (LWP 24368)] [Switching to Thread -134438912 (LWP 24368)]
Breakpoint 1, foo (x=Cannot access memory at address 0x0 ) at myprogram.o.optdbg:3 3 1| REAL*4 FUNCTION foo (x, y) (gdb) step 4 8| #2 = _xlfBeginIO(",257,#1,32768,NULL,0,NULL) (gdb) step 5 T_2 = (x + y) + 1.00000000E+00 (gdb) step 4 8| #2 = _xlfBeginIO(",257,#1,32768,NULL,0,NULL) (gdb) step 5 T_2 = (x + y) + 1.00000000E+00 (gdb) step 6 CALL _xlfWriteLDReal(%VAL(#2),T_2,",") (gdb) step 8.000000000 7 _xlfEndIO(%VAL(#2))
24 XL Fortran: Optimization and Programming Guide (gdb) step 8 #2.rnn5 = _xlfBeginIO(",257,#1,32768,NULL,0,NULL) (gdb) cont Continuing. 9.000000000 10.00000000 11.00000000 7.000000000
Program exited normally.
Tracing procedures in your code You can instruct the compiler to insert calls to the tracing procedures that you have defined to aid in debugging or timing the execution of other procedures.
To trace procedures in your program, you must specify which procedures to trace. You must also provide your own tracing procedures. If you enable tracing without providing tracing procedures, you will get linker errors about undefined symbols called __func_trace_enter, __func_trace_exit, and possibly __func_trace_catch. Specifying which procedures to trace
The -qfunctrace compiler option controls tracing for all non-inlined user-defined procedures and all outlined compiler-generated procedures in your program. If you are interested in tracing specific external or modules procedures, you can use the -qfunctrace+ and -qfunctrace- compiler options. You can also specify the NOFUNCTRACE directive to disable the tracing of entire modules, external procedures, module procedures, or internal procedures. What can be traced
Tracing applies to programs, external procedures, non-intrinsic module procedures, and internal procedures.
Compiler-generated procedures are not traced unless they were generated for outlined user code, such as an OpenMP program. In those cases, the name of the outlined procedure contains the name of the original user procedure as a prefix.
Inlined procedures and statement functions cannot be traced because they do not exist in the executable.
To avoid infinite recursion, user-defined tracing procedures cannot be traced. Similarly, tracing must be disabled for procedures called from user-defined tracing procedures. How to write tracing procedures
You can implement the tracing procedures in Fortran, C, or C++.
To implement the tracing procedures in Fortran, the characteristics of the procedures must be the same as those specified in the following interface: SUBROUTINE routine_name(procedure_name, file_name, line_number, id) USE, INTRINSIC :: iso_c_binding CHARACTER(*), INTENT(IN) :: procedure_name
Chapter 1. Optimizing your applications 25 CHARACTER(*), INTENT(IN) :: file_name INTEGER(C_INT), INTENT(IN) :: line_number TYPE(C_PTR), INTENT(INOUT) :: id END SUBROUTINE
where routine_name is the name of an external or module procedure.
You must then tell the compiler to use your subroutine as a tracing procedure in one of the following ways: v Using the -qfunctrace_xlf_enter, -qfunctrace_xlf_exit,or-qfunctrace_xlf_catch compiler options. v Using the FUNCTRACE_XLF_ENTER, FUNCTRACE_XLF_EXIT,or FUNCTRACE_XLF_CATCH directives.
When you specify these options or directives, XL Fortran generates wrapper procedures called __func_trace_enter, __func_trace_exit, and __func_trace_catch that call your corresponding tracing procedure. These wrappers allow interoperability with C and C++ by converting the dummy arguments from the C prototype to the interface described earlier. routine_name must therefore not be named __func_trace_enter, __func_trace_exit,or __func_trace_catch. In addition, your program must not contain more than one of each of the tracing procedures.
Writing the tracing procedures in C or C++ requires that you provide the __func_trace_enter, __func_trace_exit, and __func_trace_catch procedures directly. They must have the following prototypes: v void __func_trace_enter(const char *const procedure_name, const char *const file_name, int line_number, void **const id); v void __func_trace_exit(const char *const procedure_name, const char *const file_name, int line_number, void **const id); v void __func_trace_catch(const char *const procedure_name, const char *const file_name, int line_number, void **const id);
Note: If you write the tracing procedures in C++, they must be declared extern "C".
XL Fortran inserts calls to your tracing procedures on procedure entry and exit. It passes the name of the procedure being traced, the name of the file containing the entry or exit point being traced, and the line number. It also passes the address of a static pointer that is initialized to C_NULL_PTR at the beginning of the program. This pointer allows you to store arbitrary data in the entry tracing procedure and access this data in the exit and catch procedures. See the Examples section for detail. Because this pointer resides in static memory, extra steps might be needed when tracing threaded or recursive procedures. Sample tracing procedures
XL Fortran provides sample tracing procedures in the /opt/ibmcmp/xlf/bg/14.1/ samples/functrace directory. You can use these procedures for simple tracing, or you can modify them for more complex tracing. v tracing_routines.c: Provides tracing procedures written in C. This file is useful when you do not require access to Fortran modules, and when there is a possibility of recursive input / output.
26 XL Fortran: Optimization and Programming Guide v tracing_routines.f90: Provides tracing procedures written in Fortran. This file is useful when you need access to Fortran modules or intrinsics in your tracing procedures. The following example illustrates the use of the samples for simple tracing: > cat helloworld.f print *, ’hello world’ end > cc -c /opt/ibmcmp/xlf/bg/14.1/samples/functrace/tracing_routines.c > bgxlf95 helloworld.f -qfunctrace tracing_routines.o ** _main === End of Compilation 1 === 1501-510 Compilation successful for file helloworld.f. > ./a.out { _main (helloworld.f:1) hello world } _main (helloworld.f:2)
>
Tracing limitations
The procedure tracing functionality has the following limitations: v A procedure cannot be traced separately from its ENTRY points. Either all are traced or none are. The name of the procedure is passed to the tracing procedure even when tracing the ENTRY point. The line number helps distinguish what is being traced in this case. v The Fortran standard requires pure procedures to have no side effects. The compiler uses this assumption when optimizing your program. If you enable tracing of a pure procedure, your tracing procedure must not change the program state in a way that creates a side effect. v The Fortran standard imposes limits on recursive input/output. If you write your tracing procedures in Fortran, you must be careful not to break these rules. The following example has a print statement where an I/O item is the result of a function call (foo). It is illegal for the tracing procedure in this case to have I/O on an external file: > cat recursive.f integer function test() test = 1 end function
integer test print *, test() ! test must not have I/O on external unit end > bgxlf95 -c /opt/ibmcmp/xlf/bg/14.1/samples/functrace/tracing_routines.f90 ** my__func_trace_enter === End of Compilation 1 === ** my__func_trace_exit === End of Compilation 2 === ** my__func_trace_catch === End of Compilation 3 === 1501-510 Compilation successful for file tracing_routines.f90. > bgxlf95 recursive.f tracing_routines.o -qfunctrace ** test === End of Compilation 1 === ** _main === End of Compilation 2 === 1501-510 Compilation successful for file recursive.f. > ./a.out { _main (recursive.f:6) XL Fortran (I/O initialization): I/O recursion detected. Aborted >
Note: You can work around this by writing the tracing procedure in C. For an example, see the tracing_routines.c sample file described in section “Sample tracing procedures” on page 26.
Chapter 1. Optimizing your applications 27 v When optimizing your program, the compiler reorders code and removes dead code. As a result, the line number passed to the tracing procedure might not be accurate when optimization is enabled. Examples
In the following example, -qfunctrace is used to measure the time spent in each external procedure. The FUNCTRACE_XLF_ENTER and FUNCTRACE_XLF_EXIT directives are used to specify procedures my_enter and my_exit as the tracing procedures. The NOFUNCTRACE directive is used to disable tracing of main_program: > cat example.f ! Designate my_enter as a tracing procedure that should be called ! on procedure entry !ibm* functrace_xlf_enter subroutine my_enter(procedure_name, file_name, line_number, id) use, intrinsic :: iso_c_binding use, intrinsic :: xlfutility character(*), intent(in) :: procedure_name, file_name integer(c_int), intent(in) :: line_number type(c_ptr), intent(inout) :: id
integer(kind=time_size), pointer :: enter_count
! Store the time we entered the procedure being traced into id. if (.not. c_associated(id)) then allocate(enter_count) enter_count = time_() id = c_loc(enter_count) end if
print *, ’Entered procedure ’, procedure_name, ’ at ( ’, & file_name, ’ :’, line_number, ’).’ end subroutine
! Designate my_exit as a tracing procedure that should be called ! on procedure exit !ibm* functrace_xlf_exit subroutine my_exit(procedure_name, file_name, line_number, id) use, intrinsic :: iso_c_binding use, intrinsic :: xlfutility character(*), intent(in) :: procedure_name, file_name integer(c_int), intent(in) :: line_number type(c_ptr), intent(inout) :: id
integer(kind=time_size), pointer :: enter_count integer(kind=time_size) exit_count, duration
! id should have been associated in my_enter with the time we ! entered the procedure being traced. Find the elapsed time. if (c_associated(id)) then exit_count = time_() call c_f_pointer(id, enter_count) duration = exit_count - enter_count else stop "error!" endif
print *, ’Leaving procedure ’, procedure_name, ’ at ( ’, & file_name, ’ :’, line_number, ’).’ print *, ’Spent’, duration, ’seconds in ’, procedure_name, ’.’ end subroutine
! sub2 will be traced subroutine sub2
28 XL Fortran: Optimization and Programming Guide call sleep_(3) end subroutine
! sub1 will be traced subroutine sub1 call sleep_(5) call sub2 end subroutine
! Do not want to trace main_program !ibm* nofunctrace program main_program call sub1 end program > bgxlf95 example.f -qfunctrace ** my_enter === End of Compilation 1 === ** my_exit === End of Compilation 2 === ** sub2 === End of Compilation 3 === ** sub1 === End of Compilation 4 === ** main_program === End of Compilation 5 === 1501-510 Compilation successful for file example.f. > ./a.out Entered procedure sub1 at ( example.f : 59 ). Entered procedure sub2 at ( example.f : 54 ). Leaving procedure sub2 at ( example.f : 55 ). Spent 3 seconds in sub2. Leaving procedure sub1 at ( example.f : 61 ). Spent 8 seconds in sub1. >
Related information v For details about the -qfunctrace compiler option, see -qfunctrace in the XL Fortran Compiler Reference. v For details about -qfunctrace_xlf_catch, -qfunctrace_xlf_enter,or -qfunctrace_xlf_exit compiler options, see the Detailed descriptions of the XL Fortran compiler options section in the XL Fortran Compiler Reference. v For details about the FUNCTRACE_XLF_CATCH, FUNCTRACE_XLF_ENTER, and FUNCTRACE_XLF_EXIT directives, see Detailed directive descriptions section in the XL Fortran Language Reference. v For details about the NOFUNCTRACE directive, see NOFUNCTRACE in the XL Fortran Language Reference.
Getting more performance The XL compiler family offers other strategies and tuning alternatives for increasing performance.
Whether you are already optimizing at -O5, or you are looking for more opportunities to increase performance without the resource costs of optimizing at higher levels, the XL compiler family offers other strategies and tuning alternatives. For more information, see the following topics: v Tuning XL compiler applications v Advanced optimization concepts v Optimizing your SMP code
Chapter 1. Optimizing your applications 29 Beyond performance: effective programming techniques Applications that perform well begin with applications that are written well. See the following topics for information about writing better code; whether your goal is to make your code more portable, more easily optimized, or interoperable with other languages. v Chapter 4, “Managing code size,” on page 45 v Chapter 5, “Compiler-friendly programming techniques,” on page 51 v Chapter 7, “Parallel programming with XL Fortran,” on page 69 v Chapter 8, “Interlanguage calls,” on page 253
30 XL Fortran: Optimization and Programming Guide Chapter 2. Tuning XL compiler applications
Included as part of the XL Fortran optimization suite are options you can use to instruct the compiler to generate code that executes optimally on a given processor or architecture family, and to instruct the compiler on the execution characteristics of your application.
The better you can convey those characteristics, the more precisely the compiler can tune and optimize your application. This section assumes that you have already begun optimizing your application using the strategies found in Optimizing your applications.
Tuning for your target architecture By default, the compiler generates code that runs on all supported systems, though this code does not run optimally on all supported systems. By selecting options to target the appropriate architectures, you can optimize your application to suit the broadest possible selection of relevant processors, a range of processors within a given family, or a specific processor.
The compiler options in the Options for targeting your architecture table introduce how you can control optimizations affecting individual aspects of your target architecture. This section also goes into further detail on how you can use some of those options to ensure your application provides the best possible performance on those targets. Table 12. Options for targeting your architecture Option Behavior -q64 Generates code for a 64-bit addressing model (64-bit execution mode). -qarch Selects a family of processor architectures, or a specific architecture that the compiler will generate machine instructions for. -qarch=qp produces object code that runs on the Blue Gene/Q platforms. -qtune Focuses optimizations for execution on a given processor without restricting the processor architectures that your application can execute on. –qtune=qp specifies that optimizations are tuned for the Blue Gene/Q platforms. -qcache Defines a specific cache or memory geometry. Selecting a predefined optimization level like -O2 sets default vales for -qcache suboptions.
In addition to targeting the correct architecture for your application, it is important to select the right level of optimization. Combining the appropriate architecture settings with an optimization level that fits your application can vastly enhance performance. If you have not already done so, consult Optimizing your applications in addition to this section. Using -qarch Using -qarch you can select a machine architecture or a family of architectures on which you can run your application. Selecting the correct -qarch suboption is crucial to influencing chip-level optimization as the choice of -qarch suboption controls:
© Copyright IBM Corp. 1990, 2012 31 v The list of machine instructions available to the compiler when generating object code. v The characteristics and capabilities of the hardware the compiler will model when optimizing. v Optimization trade-offs and opportunities in individual instruction selection and instruction sequence selection v The default setting of the -qtune option.
Architecture selection is important at all optimization levels. Even at low optimization levels like -O0 and -O2, specifying the correct target architecture can be beneficial to performance. Specifying the correct target allows the compiler to select more efficient machine instructions and generate instruction sequences that perform best for a particular machine.
The -qarch suboptions allow you to specify individual processors or a family of processors with common instruction sets or subsets. The choice of processor gives you the flexibility of compiling your application to execute optimally on a particular machine, or to execute on a wide variety of machines while still applying as much architecture-specific optimization as possible. The less specific your choice of architecture, the fewer machine instructions available to the compiler when generating code. A less specific choice can also limit the number of hardware intrinsic functions available to your application. A more specific choice of architecture, can make available more instructions and hardware intrinsic functions. The XL Fortran Compiler Reference details the specific chip architectures and architecture families available.
When compiling your application, using a consistent or compatible -qarch setting for all files will ensure that you are getting the most from your architecture targets. If you are using -qipa link-time optimizations, the architecture setting you specify on the link step overrides the compile step setting.
You must ensure that your application executes only on machines that support your -qarch settings. Executing your application on other machines can produce incorrect results, even if your application appears to run without trapping. Using -qtune The -qtune option focuses optimizations for execution on a given processor without restricting the processor architectures that your application can execute on, generating machine instructions consistent with your -qarch architecture choice. Using -qtune also guides the optimizer in performing transformations, such as instruction scheduling, so that the resulting code executes most efficiently on your chosen -qtune architecture. The -qtune option tunes code to run on one particular processor architecture, and includes only specific processors as suboptions. The -qtune option does not support suboptions representing families of processors.
Use -qtune to specify the most common or critical processor where your application executes.
The default -qtune setting depends on the -qarch setting. If the -qarch option is set to a particular machine architecture, this limits the range of available -qtune suboptions, and the default tune setting will be compatible with the selected target processor. If -qarch option is set to a family of processors, the range of values available for -qtune expands across that family, and the default is chosen from a commonly used machine in that family. If you compile with -qtune=auto, the default for optimization levels -O4 and -O5, the compiler detects the machine
32 XL Fortran: Optimization and Programming Guide characteristics on which you are compiling, and assumes you want to tune for that type of machine. You can override this behavior by specifying -qtune after the -O4 or -O5 compiler options. Using -qcache The -qcache option allows you to instruct the optimizer on the memory cache layout of your target architecture. There are several suboptions you can specify to describe cache characteristics such as: v The types of cache available v The cache size v Cache-miss penalties The -qcache option is only effective if you understand the cache characteristics of the execution environment of your application. Before using -qcache, look at the options section of the listing file with the -qlist option to see if the current cache settings are acceptable. The settings appear in the listing when you compile with -qlistopt. If you are unsure about how to interpret this information, do not use -qcache, and allow the compiler to use default cache settings.
If you do not specify -qcache, the compiler makes cache assumptions based on your -qarch and -qtune settings. If you compile with the -qcache=auto suboption, the default at optimization levels -O4 and -O5, the compiler detects the cache characteristics of your compilation machine and tunes cache optimizations for that cache layout. If you do specify -qcache, also specify -qhot, or an option such as -O4 that implies -qhot. The optimizations that -qhot performs are designed to take advantage of your -qcache settings. Before you finish tuning Consult the following list to ensure that you are getting the most out of your target machine options. v Do not specify a -qarch option that is incompatible with your hardware. This can produce unexpected results. v Specify a -qarch setting that represents the largest common instruction set available to the machines that your application will execute on. v If you are executing your application on multiple machines, choose the -qtune suboption that aligns with the machine you expect your application to run on most frequently or where performance is most important. v If compiling with -qcache, specify -qhot as well, which can take advantage of your cache settings.
Tuning your code for Blue Gene This section describes the strategy that you can use to facilitate the automatic single-instruction-multiple-data (SIMD) capabilities in XL Fortran on Blue Gene/Q platforms.
Blue Gene/Q provides SIMD instructions, which can operate quadrate double words (256 bits) in parallel with one instruction. SIMD exploits the instruction level parallelism (ILP) and can greatly improve the performance of your program.
IBM XL Fortran for Blue Gene/Q, V14.1 is able to automatically generate SIMD instructions (auto-SIMD) for your program when the conditions for SIMD are
Chapter 2. Tuning XL compiler applications 33 satisfied. Auto-SIMD is enabled at the following optimization levels when -qsimd=auto and -qhot=level=1 is in effect. In particular, at -O3, -qhot=level=0 is implied and auto-SIMD is enabled too. v -O2 -qhot v -O3 v -O3 -qhot v -O4 -qhot v -O5 -qhot
Notes: v On Blue Gene/Q, -qsimd=auto is enabled by default at all optimization levels. v Specifying -qhot without suboptions is equivalent to -qhot=level=1.
You can always turn off auto-SIMD with -qsimd=noauto. Auto-SIMD for loops and basic blocks
When auto-SIMD is enabled, XL Fortran looks for two kinds of candidates, loops and basic blocks, and tries to group the operations on contiguous data into SIMD operations.
Example 1
This example shows a sample loop construct. DOi=0,100 a(i) = b(i) + c(i) END DO
When auto-SIMD is enabled, XL Fortran can transform this loop into the following pseudo code. Four iterations of this loop are run in parallel with SIMD instructions. VECTOR(REAL(8)) :: v1 VECTOR(REAL(8)) :: v2 VECTOR(REAL(8)) :: v3
DOi=0,100, 4 v1 = VEC_LDS(0, b(i)) ! Loads b(i), b(i+1), b(i+2), and b(i+3). v2 = VEC_LDS(0, c(i)) ! Loads c(i), c(i+1), c(i+2), and c(i+3). v3 = VEC_ADD(v1, v2) ! Adds the four elements. CALL VEC_ST(v3, 0, a(i)) ! Stores the result to a(i), a(i+1), a(i+2), and a(i+3). END DO
Example 2
This example shows a sample basic block. a(1) = b(1) + c(1) a(2) = b(2) + c(2) a(3) = b(3) + c(3) a(4) = b(4) + c(4)
When auto-SIMD is enabled, XL Fortran can transform the block into the following pseudo code. The four statements are run in parallel with SIMD instructions. VECTOR(REAL(8)) :: v1 = VEC_LDS(0, b(1)) ! Loads b(1), b(2), b(3), and b(4). VECTOR(REAL(8)) :: v2 = VEC_LDS(0, c(1)) ! Loads c(1), c(2), c(3), and c(4). VECTOR(REAL(8)) :: v3 = VEC_ADD(v1, v2) ! Adds the four elements. CALL VEC_ST(v3, 0, a(1)) ! Stores the result to a(1), a(2), a(3), and a(4).
34 XL Fortran: Optimization and Programming Guide You can use the -qreport option to display which loops are optimized by auto-SIMD and why other loops are not. Conditions for auto-SIMD transformation
In auto-SIMD transformation, XL Fortran verifies the following conditions that guarantee the correctness of the transformation: v The dependence constraint for SIMD You can help the compiler by providing the alias information , or use the INDEPENDENT directive to specify that the loop iterations are independent. !IBM* INDEPENDENT DOi=1,n a(i + l1) = a(i + l2) + a(i + l3) END DO v The alignment of data access The alignment of pointers might become unknown to the compiler when pointer assignment or pointer arithmetic are involved. You can help the compiler by providing the alignment information with the ALIGN directive, or using the -qassert=refalign option to assert that all the pointers are naturally aligned to the type of data being pointed to. The following example shows how to use the ALIGNX intrinsic procedure to assert the alignment. CALL ALIGNX(32, a) CALL ALIGNX(32, b) CALL ALIGNX(32, c)
DOn=1,200 c(n) = a(n) + b(n) END DO
Related information v -qsimd v -qhot v -qreport v -qassert v ALIGN v INDEPENDENT v ALIGNX(K,M) v Vector intrinsic procedures (IBM extension)
Further option driven tuning You can use options to convey the characteristics of your application to the compiler, tuning the optimizations that the compiler will apply. Option driven tuning is a process that can require experimentation to find the right combination of options to increase the performance of your application.
The XL compilers support many options that allow you to assert that your application will not follow certain standard language rules in some instances. The compiler assumes language standard compliance and can perform unsafe optimizations if your application is not compliant. Standards-conforming applications are more easily optimized and more portable, but when full compliance is not possible, use the appropriate options to ensure your code is optimized safely.
Chapter 2. Tuning XL compiler applications 35 For complete compiler option syntax, see the XL Fortran Compiler Reference. Options for providing application characteristics This section provides a list of options that can dictate a wide variety of characteristics about your application to the compiler including floating-point and loop behaviors. Option Description -qalias Supports several suboptions that can help the compiler analyze the characteristics of your application. For more information on aliasing, see Advanced optimization concepts. noaryovrlp Asserts that your application contains no array assignments between storage associated (overlapping) arrays. nointptr Asserts that your application does not make use of integer (Cray) pointers. nopteovrlp Asserts that your application does not contain pointee variables that refer to any data objects that are not pointee variables. Also, that your application does not contain two pointee variables that can refer to the same storage location. std Asserts that your application follows all language rules for variable aliasing. This is the default compiler setting. Specify -qalias=nostd if your application does not follow all variable aliasing rules. -qassert Includes the following suboptions that can be useful for providing some loop characteristics of your application. nodeps Asserts that the loops in your application do not contain loop carry dependencies. itercnt=number Gives the optimizer a value to use when estimating the number of iterations for loops where it cannot determine that value. -qddim Forces the compiler to reevaluate the bounds of a pointee array each time the application references the array. Specify this option only if your application performs dynamic dimensioning of pointee arrays. -qdirectstorage Asserts that your application accesses write-through-enabled or cache-inhibited storage. -qfloat Provides the compiler with floating-point characteristics for your application. The following suboptions are particularly useful. nans Asserts that your application makes use of signaling NaN (not-a-number) floating-point values. Normal floating-point operations do not create these values, your application must create signalling NaNs.
36 XL Fortran: Optimization and Programming Guide rrm Prohibits optimization transformations that assume the floating-point rounding mode must be the default setting round-to-nearest. If your application changes the rounding mode in any way, specify this option. -qflttrap Controls various aspects of floating-point exception handling that your application can require if it attempts to detect or handle such exceptions. -qieee Specifies the preferred floating-point rounding mode when evaluating expressions at compile time. This option is important if your application requires a non-default rounding mode in order to have consistency between compile-time evaluation and runtime evaluation. You can also specify -y to set the preferred floating-point rounding mode. -qlibansi Asserts that any external function calls in your compilation that have the same name as standard C library function calls, such as malloc or memcpy, are in fact those functions and are not a user-written function with that name. -qlibessl Asserts that your application will be linked with IBM's ESSL high-performance mathematical library and that mathematical operations can be transformed into calls to that library. For more information on ESSL, see the High performance libraries topic. -qlibmpi Asserts that all functions with Message Passing Interface (MPI) names are in fact MPI functions and not a user function with different semantics. -qlibposix Asserts that any external function calls in your application that have the same name as standard Posix library function calls are in fact those functions and are not a user-written function with that name. -qonetrip Asserts that all DO loops in your application will execute at least one iteration. You can also specify this behavior with -1. -qnostrictieeemod Relaxes certain rules required by the Fortran 2003 standard related to the use of the IEEE intrinsic modules. Specify this option if your application does not use these modules. -qstrict_induction Prevents optimization transformations that would be unsafe if DO loop integer iteration count variables overflow and become negative. Few applications contain algorithms that require this option. -qthreaded Informs the compiler that your application will execute in a multithreaded/SMP environment. Using an _r invocation, like bgxlf_r, adds this option automatically. -qnounwind Informs the compiler that the stack will not be unwound while any routine in your application is active. The -qnounwind option enables prologue tailoring optimization, which reduces the number of saves and restores of nonvolatile registers.
Chapter 2. Tuning XL compiler applications 37 -qnozerosize Asserts that this application does not require checking for zero-sized arrays when performing array operations. Options to control optimization transformations There are many options available to you in addition to the base set found in the Optimizing your applications section. Some of these options prevent an optimization that can be unsafe for certain applications or enable one that is safe for your application, but is not normally available as part of the optimization process. Option Description -qcompact Chooses a reduction of final code size over a reduction in execution time. You can use this option to constrain the optimizations of -O3 and higher. For more information on restriction code size, see the Managing code size section. -qsimd=auto Makes use of the vector capabilities of chips. -qfloat This option provides a number of suboptions for controlling the optimizations to your floating-point calculations. norsqrt Prevents the replacement of the division of the result of a square-root calculation with a multiplication by the reciprocal of the square root. nostrictmaf Prevents certain floating-point multiply-and-add instructions where the sign of signed zero value would not be preserved. -qipa Includes many suboptions that can assist the IPA optimizations while analyzing your application. If you are using the -qipa option or higher optimization levels that imply IPA, it is to your benefit to examine the suboptions available. -qmaxmem Limits the memory available to certain memory-intensive optimizations at low levels. Specify -qmaxmem=-1 to remove these memory limits. -qnoprefetch Prevents the the insertion of prefetching machine instructions into your application during optimization. -qinline Exerts control over inlining optimization transformations. For more information on inlining, see the Advanced optimization concepts section. -qsmallstack Instructs the compiler to limit the use of stack storage in your application. This can increase heap usage. -qsmp Produces code for an SMP system. This option also searches for opportunities to increase performance by automatically parallelizing your code. The Parallel programming with XL Fortran section contains more information on writing parallel code.
38 XL Fortran: Optimization and Programming Guide -qstacktemp Limits certain compiler temporaries allocated on the stack. Those not allocated on the stack will be allocated on the heap. This option is useful for applications that use enough stack space to exceed stack user or system limits. -qstrict Limits optimizations to strict adherence to implied program semantics. This often prevents the compiler from ignoring certain little-used rules in the IEEE floating-point specification that few applications require for correct behavior. For example, reordering or reassociating a sequence of floating-point calculations can cause floating-point exceptions at an unexpected location or mask them completely. The -qstrict option includes suboptions that refine the control of the transformations performed by the optimizers. Do not use this option unless your application requires strict adherence as -qstrict and its suboptions can severely inhibit optimization. -qunroll Independently controls loop unrolling. At -O3 and higher, -qunroll is a default setting. Options to assist with performance analysis The compiler provides a set of options that can help you analyze the performance aspects of your application. These options are most useful when you are selecting your level of optimization and tuning the optimization process to the particular characteristics of your application. -d Informs the compiler that you want to preserve the preprocessed versions of your compilation files. Typically these files would have a .F extension. -g inserts full debugging information into your object code. While the optimization process can obscure original program meaning, at least some of the information that this option produces is useful to performance analysis tools. You can also specify this behavior with -qdbg. -p Inserts appropriate profiling information into your object to code to make using tools for performance analysis possible. You can also specify this behavior with -pg. -qdpcl Prepares your object for processing by tools based on the Dynamic Probe Class Library (DPCL). -qlinedebug An option similar to -g, this option inserts only minimal debugging information into your object code such as function names and line number information. -qlist Produces a listing file containing a pseuo-assembly listing of your object code. -qlistfmt Creates a compiler report to assist with finding optimization opportunities. -qreport Inserts information in the listing file showing the transformations done by certain optimizations. -S Produces a .s file containing the assembly version of the .o file produced by the compilation.
Chapter 2. Tuning XL compiler applications 39 -qtbtable Limits the amount of debugging traceback information in object files, which reduces the size of the program. Use -qtbtable=full if you intend to analyze your application with a profiling utility. Options that can inhibit performance Some compiler options are necessary for some applications to produce correct or repeatable results. Usually, these options instruct the compiler to enforce very strict language semantics that few applications require. Others are supported by the compiler to allow compilation of code that does not conform to language standards. Avoid these options if you are trying to increase the runtime performance of your application. In cases where these options are enabled by default, you must disable them to increase performance. You can specify -qlistopt to show, in the listing file, the settings of each of these options.
The following list summarizes the options that can inhibit performance. Each option is described in the XL Fortran Compiler Reference. v -qalias=nostd v -qcompact v -qfloat=nosqrt, -qfloat=nostrictmaf, -qfloat=rrm v -qsimd=noauto v -qnoprefetch v -qnounroll v -qsmallstack v -qstacktemp=[value other than 0 or -1] v -qstrict v -qstrict_induction v -qstrictieeemod v -qunwind v -qxlf2008=checkpresence v -qzerosize v -qnoinline
40 XL Fortran: Optimization and Programming Guide Chapter 3. Advanced optimization concepts
After you apply command-line optimizations and tuning that are appropriate to your application and the constraints of your development cycle, you have opportunities to further improve the performance of your application through aliasing and inlining.
Aliasing An alias occurs when different variables point directly or indirectly to a single area of storage. Aliasing refers to assumptions made during optimization about which variables can point to or occupy the same storage area.
When an alias exists, or the potential for an alias occurs during the optimization process, pessimistic aliasing occurs. This can inhibit optimizations like dead store elimination and loop transformations on aliased variables. Also, pessimistic aliasing can generate additional loads and stores as the compiler must ensure that any changes to the variable that occur through the alias are not lost.
When aliasing occurs there is less opportunity for optimization transformations to occur on and around aliased variables than variables where no aliasing has taken place. For example, if variables A, B, and C are all aliased, any optimization must assume that a store into or a use of A is also a store or a use of B and C, even if that is not the case. Some of the highest optimization levels can improve alias analysis and remove some pessimistic aliases. However, in all cases, when it is not proven during an optimization transformation that an alias can be removed that alias must be left in place.
Where possible, avoid programming techniques that lead to pessimistic aliasing assumptions. These aliasing assumptions are the single most limiting factor to optimization transformations. The following situations can lead to pessimistic aliasing: v When you assign a pointer the address of any variable, the pointer can be aliased with globally visible variables and with static variables visible in the pointer's scope. v When you call a procedure that has dummy arguments passed by reference, aliasing occurs for variables used as actual arguments, and for global variables. v The compiler will make several worst-case aliasing assumptions concerning variables in common blocks and modules. These assumptions can inhibit optimization.
Some compiler options like -qalias can affect aliasing directly. For more information on how to tune the aliasing behavior in your application, see “Options for providing application characteristics” on page 36.
Inlining Inlining is the process of replacing a subroutine or function call at the call site with the body of the subroutine or function being called. This eliminates call-linkage overhead and can expose significant optimization opportunities.
© Copyright IBM Corp. 1990, 2012 41 For example, with inlining, the compiler can replace the subroutine parameters in the function body with the actual arguments passed. Inlining trade-offs can include code bloat and an increase in the difficulty of debugging your source code.
If your application contains many calls to small procedures, the procedure call overhead can sometimes increase the execution time of the application considerably. Specifying the -qinline compiler option can reduce this overhead. Additionally, you can use the -p or -pg options and profiling tools to determine which subprograms your application calls most frequently, and use -qinline to list their names to ensure inlining.
The -qinline option can perform inlining where the calling and called procedures are in different compilation units. This applies to optimization level -O5 only. # Let the compiler decide what to inline. bgxlf95 -O3 -qinline inline.f
# Encourage the compiler to inline particular subprograms. bgxlf95 -O3 -qinline+called_100_times:called_1000_times inline.f
Note: -qipa=inline is deprecated and no longer supported; it is replaced by -qinline. Finding the right level of inlining A common occurrence in application optimization is excessive inlining. This can actually lead to a decrease in performance because running larger programs can cause more frequent cache misses and page faults. Because the XL compilers contain safeguards to prevent excessive inlining, this can lead to situations where subprograms you want to inline are not automatically inlined when you specify -qinline.
Some common conditions that prevent -qinline from inlining particular subprograms are: v The calling and called procedures are in different compilation units. If so, you can use the -qinline option in the link step to enable cross-file inlining. This applies to optimization level -O5 only. v After inlining expands a subprogram to a particular limit, the optimizer does not inline subsequent calls to that subprogram. v Any interface errors, such as different numbers, sizes, or types of arguments or return values, can prevent inlining for a subprogram call. You can also use interface blocks for the procedures being called. v Actual or potential aliasing of dummy arguments or automatic variables can limit inlining. Consider the following cases: – There are more than 31 arguments to the procedure your application is calling. – Any automatic variables in the called procedures are involved in an EQUIVALENCE statement – The same variable argument is passed more than once in the same call. For example, CALL SUB(X,Y,X). v Some procedures that use computed GO TO statements, where any of the corresponding statement labels are also used in an ASSIGN statement.
To change the size limits that control inlining, you can specify -qinline=level=n, where n is 0 through 10. Larger values allow more inlining.
42 XL Fortran: Optimization and Programming Guide It is possible to inline C/C++ functions into Fortran programs and Fortran functions into C/C++ programs during link time optimizations. You must compile the C/C++ code using the IBM XL C/C++ compilers with -qinline and a compatible option set to that used in the IBM XL Fortran compilation.
Chapter 3. Advanced optimization concepts 43 44 XL Fortran: Optimization and Programming Guide Chapter 4. Managing code size
Code size is often not a detriment to performance for most XL compiler programmers. For some however, generating compact object code can be as important as generating efficient code.
Oversized programs can affect overall performance by creating a conflict for real storage between pages of virtual storage containing code, and pages of virtual storage containing data. On systems with a small, combined instruction and data cache, cache collisions between code and data can also reduce performance. This section provides suggestions on how to achieve a balance between code efficiency and object-module size, while identifying compiler options that can affect object-module size. Code size tuning is most effective once you have built a stable application and run optimization at -O2 or higher.
Reasons for tuning for code size include: v Your application design calls for an implementation with limited real memory, instruction-cache space, or disk space. v When loading your application, it uses enough memory to create a conflict between code areas and data areas in real memory, and both code and data are frequently paged in and out. v There are high activity areas in your code large enough that instruction cache and instruction Translation Lookaside Buffer (TLB) misses have a major effect on performance. v You intend your application to run on a host that serves end users, or in a batch environment with limits on real memory.
Before tuning for code size, it is important for you to determine whether code size is the actual problem. Very large applications tend to have small clusters of high activity and large sections of infrequently accessed code. If a particular code page is not accessed in a particular run, that page is never loaded into memory, and has no negative impact on performance. If you are tuning for code size due to the high activity code segments that cause instruction cache and instruction TLB misses that have a major effect on performance, this can be symptomatic of a program structure that requires improvement or hardware not suited to the resource requirements of the application.
If your data takes up more real storage than is available, reducing code size can improve performance by ensuring that fewer pages of data are paged out as code is paged in. However, data blocking strategies are likely to prove both more effective and easier to implement. Processing data in each page as completely as possible before moving on to the next page can reduce the number of data page misses.
If you are coding an application for a machine with a combined instruction and data cache, you can improve performance by applying the techniques described later in this section, but tuning for data cache management can yield better results than code-size tuning. Also note that highly tuning your code for the cache characteristics of one system can lead to undesirable performance results if you execute your application elsewhere.
© Copyright IBM Corp. 1990, 2012 45 Steps for reducing code size Reducing the code size of your application can have a positive effect on the performance of your application
Consider the following steps for reducing code size: v Ensure that you have built a stable application that compiles at -O2 or higher. v Use performance analysis tools to isolate high activity code segments and tune for performance where appropriate. Basing decisions for code size tuning on an application that has already undergone performance analysis will give you more information on where your application could benefit from code size tuning. v Use compiler options like -qcompact to help reduce code size. See Compiler option influences on code size for more information. Also see the following options in the XL Fortran Compiler Reference: – -qinline. – The partition parameter for -qipa. – -qunroll.
Be aware that optimization can cause code to expand significantly through loop unrolling, invariant IF floating, inlining, and other optimizations. The higher your optimization level, the more code size can increase. For more information on finding an optimization level appropriate for your application, see Chapter 1, “Optimizing your applications,” on page 1.
Compiler option influences on code size High optimization levels can increase code size. You can use other compiler options to influence the size of your code and improve performance. The -qipa compiler option The -qipa option enables interprocedural analysis (IPA) by the compiler. Interprocedural analysis analyzes the relationships between procedures and the code that references those procedures, so that more optimizations within procedures and across procedure references can take place. Interprocedural analysis can decrease code size and improve performance at the same time. In some cases however, IPA inlining can increase code size. Use with discretion. Related reference:
See interprocedural analysis (IPA) in the Compiler Reference The -qinline inlining option Using the -qinline compiler option, you can specify that the compiler consider all Fortran 90 or Fortran 95 procedures, or a particular list of procedures for inlining. Inlining procedures can increase the performance of your application. However, if your program references a procedure from many different locations in the source code, inlining that procedure can increase code size dramatically. You can use -qnoinline to disable procedure inlining entirely. You can also partially disable inlining with -qinline-procedure_name.
Do not assume that all inlining increases code size. When your source code references a very small procedure many times, inlining can reduce code size, because inlining eliminates control transfer and data interface code. In addition, inlining code facilitates other optimizations at the point of inlining, by providing
46 XL Fortran: Optimization and Programming Guide information on the values of arguments referencing the procedure. If a procedure is very small and is referenced from a number of places, inlining can also increase code locality and reduce code paging.
For details about the -qinline compiler option, see -qinline in the XL Fortran Compiler Reference. The -qhot compiler option The loop analysis and optimization available when you specify -qhot can increase code size. If your application contains many large loops and loop optimization opportunities exist, -qhot code size can increase significantly along with performance. Specifying -qhot=level=0 will perform minimal high-order transformations if code size is an issue. The topic High-order transformation contains more information on using -qhot effectively. The -qcompact compiler option The -qcompact compiler option instructs the compiler to avoid certain optimizing transformations that expand the object code. Compiling with -qcompact, disables many transformations, including: v Loop unrolling v Expansion of fixed-point multiply by more than one instruction v Inline expansion of some string and memory manipulation functions. In some cases -qcompact will avoid inlining opportunities entirely.
Specifying -qcompact creates a trade-off between the performance of individual routines in your application, and overall system performance. Suppressing transformations degrades the performance of individual routines, while overall system performance can increase as a more compact program can provide some or all of the following: v Fewer instruction-cache misses v Fewer TLB misses for pages of application code v Fewer page faults for application code
Other influences on code size In addition to compiler options, there are a number of ways programming and analysis can influence the size of your source code. High activity areas Once you apply the techniques discussed earlier in this section, your strategy for further code size reduction depends on your objective. Use profiling tools to locate hot spots in your program; then follow one of the following guidelines: v If you want to reduce code size to reduce program paging, concentrate on minimizing branches and procedure references within those hot spots. v If you want to reduce code size to reduce the size of your program's files on disk, concentrate on areas that are not hot spots. Remove any expansive optimizations from code that does not contain hot spots. Computed GOTOs and CASE constructs A sparse computed GOTO can increase code size considerably. In a sparse computed GOTO, most statement labels point to the default. Consider the following example where label 10 is the default:
Chapter 4. Managing code size 47 GOTO (10,10,10,10,20,10,10,10,10,30,20,10,10,10,10, +10,20,10,20,10,20,30,30,10,10,10,10,10,10,20,10,10,... +10,20,30,10,10,10,30,10,10,10,10,10,10,10,20,10,30) IA(I)
GOTO 10 30 CONTINUE ! ... GOTO 10 20 CONTINUE ! ... 10 CONTINUE
Although fewer cases are shown, the following CASE construct is a functionally equivalent to the example above. N is the value of the largest integer that the computed GOTO or CASE construct is testing. INTEGER IA(10000) SELECT CASE (IA(I)) CASE DEFAULT GOTO 10 CASE (5) GOTO 20 CASE (10) GOTO 30 CASE (11) GOTO 20 ! ... CASE (N-10) GOTO 30 CASE (N-2) GOTO 20 CASE (N) GOTO 30 END SELECT
In both examples, the compiler builds a branch table in the object file that contains one entry for each possibility from 1 to N, where N is the largest integer value tested. The data section of the program stores this branch table. If N is very large, the table can increase both the size of the object file and the effects of data-cache misses.
If you use a CASE construct with a small number of cases and wide gaps between the test values of the cases, the compiler selects a different algorithm to dispatch to the appropriate location, and the resulting code can be more compact than a functionally equivalent computed GOTO. The compiler cannot determine that a computed GOTO has a default branch point, so the compiler assumes that any value in the range will be selected. In a CASE construct, the compiler assumes that cases you do not specify in the construct are handled as default. Code size with dynamic or static linking Dynamic or static linking each affect the size of your code, and the resulting performance of your application. Dynamic linking and code size
When linking your programs, dynamic linking often ensures more compact code than linking statically. Dynamic linking does not include library procedures in your object file. Instead, a reference at runtime causes the operating system to locate the dynamic library that contains the procedure, and reference that procedure from the library on the system. Only one copy of the procedure is in memory, even if several programs, or copies of a single program, are accessing the procedure
48 XL Fortran: Optimization and Programming Guide simultaneously. This can reduce paging overhead. However, any libraries your program references must be present in your application's execution environment.
Note that if your program references high performance libraries like BLAS or ESSL, these procedures are dynamically linked to your program by default. Static linking and code size
Static linking binds library procedures into your application's object file. This can increase the size of your object file. If your program references only a small portion of the procedures available in a library, static linking can eliminate the need to provide the library to your users. However, static linking ties your application to one version of the library which can be detrimental in situations where your application will execute in different environments, such as different levels of the operating system.
Chapter 4. Managing code size 49 50 XL Fortran: Optimization and Programming Guide Chapter 5. Compiler-friendly programming techniques
Writing compiler-friendly code, with both the optimizer and portability in mind, can be as important to the performance of your application as the compilation options that you specify.
General practices It is not necessary to hand-optimize your code, as hand-optimizing can introduce unusual constructs that can obscure the intentions of your application from the compiler and limit optimization opportunities.
Large programs can use significant address space resources.
Avoid breaking your program into too many small functions, as this can increase the percentage of time the program spends in dealing with call overhead. If you choose to use many small functions, compiling with -qipa can help minimize the impact on performance. Attempting to optimize an application with many small functions without the benefit of -qipa can severely limit the scope of other optimizations.
Use command invocations like bgxlf90 and bgxlf95, which use -qnosave. The -qnosave option sets the default storage class of all variables to automatic. This provides more opportunities for optimization. All compiler command invocations except bgf77, bgfort77, bgxlf and bgxlf_r use -qnosave by default.
Use modules to group related subroutines and functions.
Use module variables instead of common blocks for global storage.
Mark all code that accesses or manipulates data objects by independent I/O processes and independent, asynchronously interrupting processes as VOLATILE. For example, mark code that accesses shared variables and pointers to shared variables. Mark your code carefully however, as VOLATILE is a barrier to optimization as accessing a VOLATILE object forces the compiler to always load the value from storage. This prevents powerful optimizations such as constant propagation or invariant code motion.
The XL compilers support high performance libraries that can provide significant advantages over custom implementations or generic libraries.
Variables and pointers The effective use of aliasing and of variables and pointers provides opportunities for improved performance and further optimization.
Obey all aliasing rules. Avoid specifying -qalias=nostd. For more information on aliasing and how it can affect performance, see “Aliasing” on page 41.
Avoid unnecessary use of global variables and pointers, including module variables and common blocks. When using global variables and pointers in a loop, load them into a local variable before the loop and store them back after. If you do not use the local variable somewhere other than in the loop body, the optimization
© Copyright IBM Corp. 1990, 2012 51 process can usually recognize what you are doing and expose more optimization opportunities. Replacing a global variable in a loop with a local variable reduces the possibilities for aliasing.
Use the INTENT statement to describe the usage of dummy arguments.
Limit the use of ALLOCATABLE objects and POINTER variables to situations demanding dynamic memory allocation.
Arrays Where possible, use local variables instead of global variables for loop index variables and bounds.
Whenever possible, ensure references to arrays or array sections refer to contiguous blocks of storage. Noncontiguous memory array references, when passed as parameters, lead to copy-in and copy-out operations.
F2008 When declaring an array pointer or an assumed-shape array, you can use the CONTIGUOUS attribute to ensure that the array elements in order are stored in contiguous memory and not separated by other data objects. An array pointer with the CONTIGUOUS attribute can only be pointer associated with a contiguous target. An assumed-shape array with the CONTIGUOUS attribute is always contiguous; however, the corresponding actual argument can be contiguous or noncontiguous. If it is noncontiguous, the compiler makes it contiguous by creating a temporary contiguous argument. When the CONTIGUOUS attribute is used, the compiler can perform appropriate semantic check and detect invalid codes, which helps you write more optimized codes and enables the compiler to further optimize the runtime performance and storage layout. F2008
Keep your array expressions simple so that the optimizer can deduce access patterns more easily and reuse index calculations in whole or in part.
Frequent use of array-to-array assignment and WHERE constructs can impact performance by increasing temporary storage and creating loops. Using -qlist or -qreport can help you understand the performance characteristics of your code, and where applying -qhot could be beneficial. If you are already optimizing with -qipa, ensure you are using the list=filename option, so that the -qlist listing file is not overwritten. Related information v F2008 The CONTIGUOUS attribute F2008
Choosing appropriate variable sizes Improve the efficiency of your application by choosing the appropriate variable sizes.
When programming SMP applications, use the CONTAINS statement only to share thread local storage.
In most cases using INTEGER(8) for scalars improves the efficiency of DO loops, subscripting, mathematical calculations and calling conventions when passing objects. However, if your code contains large arrays with values that can fit in an INTEGER(4), using smaller kind parameters can actually improve memory efficiency by reducing memory traffic to load or store data.
52 XL Fortran: Optimization and Programming Guide Use the lowest floating-point precision appropriate to your application. Higher precisions can reduce performance, so use the REAL(16),orCOMPLEX(16) data types only when you require extremely high precision.
On systems with QPX, -qsimd=auto provides opportunities for vectorization on REAL(8) types.
Chapter 5. Compiler-friendly programming techniques 53 54 XL Fortran: Optimization and Programming Guide Chapter 6. High performance libraries
XL Fortran is shipped with a set of libraries for high-performance mathematical computing.
The set of libraries for high-performance mathematical computing are: v The Mathematical Acceleration Subsystem (MASS) is a set of libraries of tuned mathematical intrinsic routines that provide improved performance over the corresponding standard system math library routines. MASS is described in “Using the Mathematical Acceleration Subsystem libraries (MASS).” v The Basic Linear Algebra Subprograms (BLAS) are a subset of routines from IBM's Engineering and Scientific Subroutine Library (ESSL) library, which provides matrix/vector multiplication functions tuned for Blue Gene architectures. The BLAS functions are described in “Using the Basic Linear Algebra Subprograms – BLAS” on page 65.
Note that if you are going to link your application with the ESSL libraries, using -qessl and IPA allows the optimizer to automatically use ESSL routines.
Using the Mathematical Acceleration Subsystem libraries (MASS) XL Fortran is shipped with a set of Mathematical Acceleration Subsystem (MASS) libraries for high-performance mathematical computing.
The MASS libraries consist of a library of scalar Fortran routines described in “Using the scalar library” on page 56, a set of vector libraries tuned for specific architectures described in “Using the vector libraries” on page 58, and a SIMD library described in “Using the SIMD library” on page 62. The functions contained in both scalar and vector libraries are automatically called at certain levels of optimization, but you can also call them explicitly in your programs. Note that the accuracy and exception handling might not be identical in MASS functions and system library functions.
The MASS functions must run with the default rounding mode and floating-point exception trapping settings.
When you compile programs with any of the following sets of options: v -qhot -qnostrict v -qhot -O3 v -O4 v -O5 the compiler automatically attempts to vectorize calls to system math functions by calling the equivalent MASS vector functions (with the exceptions of functions vatan2, vsatan2, vdnint, vdint, vcosisin, vscosisin, vqdrt, vsqdrt, vrqdrt, vsrqdrt, vpopcnt4, vpopcnt8, vexp2, vexp2m1, vsexp2, vsexp2m1, vlog2, vlog21p, vslog2, and vslog21p). If it cannot vectorize, it automatically tries to call the equivalent MASS scalar functions. For automatic vectorization or scalarization, the compiler uses versions of the MASS functions contained in the XLOPT library libxlopt.a.
© Copyright IBM Corp. 1990, 2012 55 In addition to any of the preceding sets of options, when the -qipa option is in effect, if the compiler cannot vectorize, it tries to inline the MASS scalar functions before deciding to call them.
“Compiling and linking a program with MASS” on page 64 describes how to compile and link a program that uses the MASS libraries, and how to selectively use the MASS scalar library functions in conjunction with the regular system libraries. Related external information Mathematical Acceleration Subsystem website, available at http://www.ibm.com/software/awdtools/mass/ Using the scalar library The MASS scalar library libmass.a contains an accelerated set of frequently used math intrinsic functions that provide improved performance over the corresponding standard system library functions. The MASS scalar functions are used when explicitly linking libmass.a.
If you want to explicitly call the MASS scalar functions, you can take the following steps: 1. Link the MASS scalar library libmass.a with your application. For instructions, see “Compiling and linking a program with MASS” on page 64 2. All the MASS scalar routines, except those listed in step 3 are recognized by XL Fortran as intrinsic functions, so no explicit interface block is needed. To provide an interface block for the functions listed in step 3, include mass.include in your source file. 3. Include mass.include in your source file for the following functions: v acosf, acosh, acoshf, asinf, asinh, asinhf, atan2f, atanf, atanh, atanhf, cbrt, cbrtf, copysign, copysignf, cosf, coshf, cosisin, erff, erfcf, expf, expm1f, hypot, hypotf, lgammaf, logf, log10f, log1pf, rsqrt, sinf, sincos, sinhf, tanf, tanhf, and x**y
The MASS scalar functions accept double-precision parameters and return a double-precision result, or accept single-precision parameters and return a single-precision result, except sincos which gives 2 double-precision results. They are summarized in Table 13. Table 13. MASS scalar functions Double- Single- Arguments Description precision precision function function acos acosf (x) Returns the arccosine of x acosh acoshf (x) Returns the hyperbolic arccosine of x anint (x) Returns the rounded integer value of x asin asinf (x) Returns the arcsine of x asinh asinhf (x) Returns the hyperbolic arcsine of x atan2 atan2f (x,y) Returns the arctangent of x/y atan atanf (x) Returns the arctangent of x atanh atanhf (x) Returns the hyperbolic arctangent of x cbrt cbrtf (x) Returns the cube root of x
56 XL Fortran: Optimization and Programming Guide Table 13. MASS scalar functions (continued) Double- Single- Arguments Description precision precision function function copysign copysignf (x,y) Returns x with the sign of y cos cosf (x) Returns the cosine of x cosh coshf (x) Returns the hyperbolic cosine of x cosisin (x) Returns a complex number with the real part the cosine of x and the imaginary part the sine of x. dnint (x) Returns the nearest integer to x (as a double) erf erff (x) Returns the error function of x erfc erfcf (x) Returns the complementary error function of x exp expf (x) Returns the exponential function of x expm1 expm1f (x) Returns (the exponential function of x)-1 hypot hypotf (x,y) Returns the square root of x2 +y2 lgamma lgammaf (x) Returns the natural logarithm of the absolute value of the Gamma function of x log logf (x) Returns the natural logarithm of x log10 log10f (x) Returns the base 10 logarithm of x log1p log1pf (x) Returns the natural logarithm of (x +1) rsqrt (x) Returns the reciprocal of the square root of x sin sinf (x) Returns the sine of x sincos (x,s,c) Sets s to the sine of x and c to the cosine of x sinh sinhf (x) Returns the hyperbolic sine of x sqrt (x) Returns the square root of x tan tanf (x) Returns the tangent of x tanh tanhf (x) Returns the hyperbolic tangent of x x**y (x,y) Returns x raised to the power y
The following example shows the XL Fortran interface declaration for the rsqrt scalar function: interface
real*8 function rsqrt (%val(x)) real*8 x ! Returns the reciprocal of the square root of x. end function rsqrt
end interface
Notes: v The trigonometric functions (sin, cos, tan) return NaN (Not-a-Number) for large arguments (where the absolute value is greater than 250pi).
Chapter 6. High performance libraries 57 v In some cases, the MASS functions are not as accurate as the libm.a library, and they might handle edge cases differently (sqrt(Inf), for example). v See the Mathematical Acceleration Subsystem website for accuracy comparisons with libm.a. Related external information Mathematical Acceleration Subsystem website, available at http://www.ibm.com/software/awdtools/mass/ Using the vector libraries If you want to explicitly call any of the MASS vector functions, you can do so by including massv.include in your source files and linking your application with the appropriate vector library. (Information about linking is provided in “Compiling and linking a program with MASS” on page 64.) libmassv.a Contains functions that have been tuned for the Blue Gene/Q architecture.
The single-precision and double-precision floating-point functions contained in the vector libraries are summarized in Table 14 on page 59. The integer functions contained in the vector libraries are summarized in Table 15 on page 60.
With the exception of a few functions (described in the following paragraph), all of the floating-point functions in the vector libraries accept three arguments: v A double-precision (for double-precision functions) or single-precision (for single-precision functions) vector output argument. v A double-precision (for double-precision functions) or single-precision (for single-precision functions) vector input argument. v An integer vector-length argument. The functions are of the form function_name (y,x,n)
where y is the target vector, x is the source vector, and n is the vector length. The arguments y and x are assumed to be double-precision for functions with the prefix v, and single-precision for functions with the prefix vs. As an example, the following code: include ’massv.include’
real*8 x(500), y(500) integer n n = 500 ... call vexp (y, x, n)
outputs a vector y of length 500 whose elements are exp(x(i)), where i=1,...,500.
The functions vdiv, vsincos, vpow, and vatan2 (and their single-precision versions, vsdiv, vssincos, vspow, and vsatan2) take four arguments. The functions vdiv, vpow, and vatan2 take the arguments (z,x,y,n). The function vdiv outputs a vector z whose elements are x(i)/y(i), where i=1,...,n. The function vpow outputs a vector z whose elements are x(i)y(i), where i=1,..,n. The function vatan2 outputs a vector z whose elements are atan(x(i)/y(i)), where i=1,..,n. The function vsincos takes the arguments (y,z,x,n), and outputs two vectors, y and z, whose elements are sin(x(i)) and cos(x(i)), respectively.
58 XL Fortran: Optimization and Programming Guide In vcosisin(y,x,n) and vscosisin(y,x,n), x is a vector of n elements and the function outputs a vector y of n complex(8)(for vcosisin) or complex(4)(for vscosisin) elements of the form (cos(x(i)),sin(x(i))). Table 14. MASS floating-point vector library functions Double-precision Single-precision function function Arguments Description vacos vsacos (y,x,n) Sets y(i) to the arc cosine of x(i), for i=1,..,n vacosh vsacosh (y,x,n) Sets y(i) to the hyperbolic arc cosine of x(i), for i=1,..,n vasin vsasin (y,x,n) Sets y(i) to the arc sine of x(i), for i=1,..,n vasinh vsasinh (y,x,n) Sets y(i) to the arc hyperbolic sine of x(i), for i=1,..,n vatan2 vsatan2 (z,x,y,n) Sets z(i) to the arc tangent of x(i)/y(i), for i=1,..,n vatanh vsatanh (y,x,n) Sets y(i) to the arc hyperbolic tangent of x(i), for i=1,..,n vcbrt vscbrt (y,x,n) Sets y(i) to the cube root of x(i), for i=1,..,n vcos vscos (y,x,n) Sets y(i) to the cosine of x(i), for i=1,..,n vcosh vscosh (y,x,n) Sets y(i) to the hyperbolic cosine of x(i), for i=1,..,n vcosisin vscosisin (y,x,n) Sets the real part of y(i) to the cosine of x(i) and the imaginary part of y(i) to the sine of x(i), for i=1,..,n vdint (y,x,n) Sets y(i) to the integer truncation of x(i), for i=1,..,n vdiv vsdiv (z,x,y,n) Sets z(i) to x(i)/y(i), for i=1,..,n vdiv_fast (1) vsdiv_fast (2) (z,x,y,n) Sets z(i) to x(i)/y(i), for i=1,..,n. vdnint (y,x,n) Sets y(i) to the nearest integer to x(i), for i=1,..,n verf vserf (y,x,n) Sets y(i) to the error function of x(i), for i=1,..,n verfc vserfc (y,x,n) Sets y(i) to the complimentary error function of x(i), for i=1,..,n vexp vsexp (y,x,n) Sets y(i) to the exponential function of x(i), for i=1,..,n vexp2 vsexp2 (y,x,n) Sets y(i) to 2 raised to the power of x(i), for i=1,..,n vexpm1 vsexpm1 (y,x,n) Sets y(i) to (the exponential function of x(i)) -1, for i=1,..,n vexp2m1 vsexp2m1 (y,x,n) Sets y(i) to (2 raised to the power of x(i)) -1, for i=1,..,n vhypot vshypot (z,x,y,n) Sets z(i) to the square root of the sum of the squares of x(i) and y(i), for i=1,..,n vlog vslog (y,x,n) Sets y(i) to the natural logarithm of x(i), for i=1,..,n vlog2 vslog2 (y,x,n) Sets y(i) to the base-2 logarithm of x(i), for i=1,..,n vlog10 vslog10 (y,x,n) Sets y(i) to the base-10 logarithm of x(i), for i=1,..,n vlog1p vslog1p (y,x,n) Sets y(i) to the natural logarithm of (x(i)+1), for i=1,..,n vlog21p vslog21p (y,x,n) Sets y(i) to the base-2 logarithm of (x(i)+1), for i=1,..,n vpow vspow (z,x,y,n) Sets z(i) to x(i) raised to the power y(i), for i=1,..,n vqdrt vsqdrt (y,x,n) Sets y(i) to the 4th root of x(i), for i=1,..,n vrcbrt vsrcbrt (y,x,n) Sets y(i) to the reciprocal of the cube root of x(i), for i=1,..,n vrec vsrec (y,x,n) Sets y(i) to the reciprocal of x(i), for i=1,..,n
Chapter 6. High performance libraries 59 Table 14. MASS floating-point vector library functions (continued) Double-precision Single-precision function function Arguments Description vrec_fast (3) vsrec_fast (4) (y,x,n) Sets y(i) to the reciprocal of x(i), for i=1,..,n vrqdrt vsrqdrt (y,x,n) Sets y(i) to the reciprocal of the 4th root of x(i), for i=1,..,n vrsqrt vsrsqrt (y,x,n) Sets y(i) to the reciprocal of the square root of x(i), for i=1,..,n vsin vssin (y,x,n) Sets y(i) to the sine of x(i), for i=1,..,n vsincos vssincos (y,z,x,n) Sets y(i) to the sine of x(i) and z(i) to the cosine of x(i), for i=1,..,n vsinh vssinh (y,x,n) Sets y(i) to the hyperbolic sine of x(i), for i=1,..,n vsqrt vssqrt (y,x,n) Sets y(i) to the square root of x(i), for i=1,..,n vtan vstan (y,x,n) Sets y(i) to the tangent of x(i), for i=1,..,n vtanh vstanh (y,x,n) Sets y(i) to the hyperbolic tangent of x(i), for i=1,..,n
Notes: 1. vdiv_fast arguments must satisfy all the following conditions for i=1,...,n: v 2-1021 ≤ |y(i)| ≤ 21020 v If x(i) is not zero, 2-969 ≤ |x(i)| < ∞ and 2-1020 ≤ |x(i)/y(i)| ≤ 21022 2. vsdiv_fast arguments must satisfy all the following conditions for i=1,...,n: v 2-125 ≤ |y(i)| ≤ 2124 v If x(i) is not zero, 2-102 ≤ |x(i)| < ∞ and 2-124 ≤ |x(i)/y(i)| ≤ 2126 3. vrec_fast arguments must satisfy the following condition for i=1,...,n: v 2-1021 ≤ |x(i)| ≤ 21020 4. vsrec_fast arguments must satisfy the following condition for i=1,...,n: v 2-125 ≤ |x(i)| ≤ 2124
Integer functions are of the form function_name (x, n), where x is a vector of 4-byte (for vpopcnt4) or 8-byte (for vpopcnt8) numeric objects (integer or floating-point), and n is the vector length. Table 15. MASS integer vector library functions Function Description Interface vpopcnt4 Returns the total number of 1 bits in the concatenation of integer*4 function vpopcnt4 (x, n) the binary representation of x(i), for i=1,...,n, where x is integer*4 x(*), n vector of 32-bit objects vpopcnt8 Returns the total number of 1 bits in the concatenation of integer*4 function vpopcnt8 (x, n) the binary representation of x(i), for i=1,...,n, where x is integer*8 x(*) vector of 64-bit objects integer*4 n
The following example shows XL Fortran interface declarations for some of the MASS double-precision vector routines: interface
subroutine vsqrt (y, x, n) real*8 y(*), x(*) integer n ! Sets y(i) to the square root of x(i), for i=1,..,n end subroutine vsqrt
60 XL Fortran: Optimization and Programming Guide subroutine vrsqrt (y, x, n) real*8 y(*), x(*) integer n ! Sets y(i) to the reciprocal of the square root of x(i), ! for i=1,..,n end subroutine vrsqrt end interface
The following example shows XL Fortran interface declarations for some of the MASS single-precision vector routines: interface subroutine vssqrt (y, x, n) real*4 y(*), x(*) integer n ! Sets y(i) to the square root of x(i), for i=1,..,n end subroutine vssqrt subroutine vsrsqrt (y, x, n) real*4 y(*), x(*) integer n ! Sets y(i) to the reciprocal of the square root of x(i), ! for i=1,..,n end subroutine vsrsqrt end interface
Overlap of input and output vectors
In most applications, the MASS vector functions are called with disjoint input and output vectors; that is, the two vectors do not overlap in memory. Another common usage scenario is to call them with the same vector for both input and output parameters (for example, vsin (y, y, n)). Other kinds of overlap (where input and output vectors are neither disjoint nor identical) should be avoided, since they may produce unexpected results: v For calls to vector functions that take one input and one output vector (for example, vsin (y, x, n)): The vectors x(1:n) and y(1:n) must be either disjoint or identical, or unexpected results may be obtained. v For calls to vector functions that take two input vectors (for example, vatan2 (y, x1, x2, n)): The previous restriction applies to both pairs of vectors y,x1 and y,x2. That is, y(1:n) and x1(1:n) must be either disjoint or identical; and y(1:n) and x2(1:n) must be either disjoint or identical. v For calls to vector functions that take two output vectors (for example, vsincos (y1, y2, x, n)): The above restriction applies to both pairs of vectors y1,x and y2,x. That is, y1(1:n) and x(1:n) must be either disjoint or identical; and y2(1:n) and x(1:n) must be either disjoint or identical. Also, the vectors y1(1:n) and y2(1:n) must be disjoint. Alignment of input and output vectors
To get the best performance from the vector library, align the input and output vectors as follows: v 16-byte for single precision v 32-byte for double precision
Chapter 6. High performance libraries 61 Consistency of MASS vector functions
All the functions in the MASS vector libraries are consistent, in the sense that a given input value will always produce the same result, regardless of its position in the vector, and regardless of the vector length. Related external information Mathematical Acceleration Subsystem website, available at http://www.ibm.com/software/awdtools/mass/ Using the SIMD library The MASS SIMD library libmass_simd.a contains a set of frequently used math intrinsic functions that provide improved performance over the corresponding standard system library functions. If you want to use the MASS SIMD functions, you can do so as follows: 1. Provide the interfaces for the functions by including mass_simd.include in your source files. 2. Link the MASS SIMD library libmass_simd.a with your application. For instructions, see “Compiling and linking a program with MASS” on page 64.
The single/double-precision MASS SIMD functions accept single/double-precision arguments and return single/double-precision results. They are summarized in Table 16. Table 16. MASS SIMD functions Double-precision Single-precision Description Double-precision function interface Single-precision function interface function function acosd4 acosf4 Computes the arc cosine of each vector(real(8)) function acosd4(vx) vector(real(8)) function acosf4(vx) element of vx. vector(real(8)), value :: vx vector(real(8)), value :: vx acoshd4 acoshf4 Computes the arc hyperbolic vector(real(8)) function acoshd4(vx) vector(real(8)) function acoshf4(vx) cosine of each element of vx. vector(real(8)), value :: vx vector(real(8)), value :: vx asind4 asinf4 Computes the arc sine of each vector(real(8)) function asind4(vx) vector(real(8)) function asinf4(vx) element of vx. vector(real(8)), value :: vx vector(real(8)), value :: vx asinhd4 asinhf4 Computes the arc hyperbolic vector(real(8)) function asinhd4(vx) vector(real(8)) function asinhf4(vx) sine of each element of vx. vector(real(8)), value :: vx vector(real(8)), value :: vx atand4 atanf4 Computes the arc tangent of vector(real(8)) function atand4(vx) vector(real(8)) function atanf4(vx) each element of vx. vector(real(8)), value :: vx vector(real(8)), value :: vx atan2d4 atan2f4 Computes the arc tangent of vector(real(8)) function atan2d4(vx,vy) vector(real(8)) function atan2f4(vx,vy) each element of vx/vy. vector(real(8)), value :: vx, vy vector(real(8)), value :: vx, vy atanhd4 atanhf4 Computes the arc hyperbolic vector(real(8)) function atanhd4(vx) vector(real(8)) function atanhf4(vx) tangent of each element of vx. vector(real(8)), value :: vx vector(real(8)), value :: vx cbrtd4 cbrtf4 Computes the cube root of each vector(real(8)) function cbrtd4(vx) vector(real(8)) function cbrtf4(vx) element of vx vector(real(8)), value :: vx vector(real(8)), value :: vx cosd4 cosf4 Computes the cosine of each vector(real(8)) function cosd4(vx) vector(real(8)) function cosf4(vx) element of vx. vector(real(8)), value :: vx vector(real(8)), value :: vx coshd4 coshf4 Computes the hyperbolic cosine vector(real(8)) function coshd4(vx) vector(real(8)) function coshf4(vx) of each element of vx. vector(real(8)), value :: vx vector(real(8)), value :: vx cosisind4 cosisinf4 Computes the cosine and sine of subroutine cosisind4 (x, y, z) subroutine cosisinf4 (x, y, z) each element of x, and stores vector(real(8)), value :: x vector(real(8)), value :: x the results in y and z as follows: vector(real(8)) y, z vector(real(8)) y, z
Sets the elements of y to cos(x1), sin(x1), cos(x2), sin(x2), and the elements of z to cos(x3), sin(x3), cos(x4), sin(x4), where x1, x2, x3, x4 are the elements of x. divd4 divf4 Computes the quotient vx/vy. vector(real(8)) function divd4(vx, vy) vector(real(8)) function divf4(vx, vy) vector(real(8)), value :: vx, vy vector(real(8)), value :: vx, vy div_fastd4 (1) div_fastf4 (2) Computes the quotient vx/vy. vector(real(8)) function div_fastd4(vx, vy) vector(real(8)) function div_fastf4(vx, vy) vector(real(8)), value :: vx, vy vector(real(8)), value :: vx, vy erfcd4 erfcf4 Computes the complementary vector(real(8)) function erfcd4(vx) vector(real(8)) function erfcf4(vx) error function of each element vector(real(8)), value :: vx vector(real(8)), value :: vx of vx. erfd4 erff4 Computes the error function of vector(real(8)) function erfd4(vx) vector(real(8)) function erff4(vx) each element of vx. vector(real(8)), value :: vx vector(real(8)), value :: vx
62 XL Fortran: Optimization and Programming Guide Table 16. MASS SIMD functions (continued) Double-precision Single-precision Description Double-precision function interface Single-precision function interface function function expd4 expf4 Computes the exponential vector(real(8)) function expd4(vx) vector(real(8)) function expf4(vx) function of each element of vx. vector(real(8)), value :: vx vector(real(8)), value :: vx exp2d4 exp2f4 Computes 2 raised to the power vector(real(8)) function exp2d4(vx) vector(real(8)) function exp2f4(vx) of each element of vx. vector(real(8)), value :: vx vector(real(8)), value :: vx expm1d4 expm1f4 Computes (the exponential vector(real(8)) function expm1d4(vx) vector(real(8)) function exp2m1f4(vx) function of each element of vx)- vector(real(8)), value :: vx vector(real(8)), value :: vx 1. exp2m1d4 exp2m1f4 Computes (2 raised to the vector(real(8)) function exp2m1d4(vx) vector(real(8)) function exp2m1f4(vx) power of each element of vx)- vector(real(8)), value :: vx vector(real(8)), value :: vx 1. hypotd4 hypotf4 For each element of vx and the vector(real(8)) function hypotd4(vx,vy) vector(real(8)) function hypotf4(vx,vy) corresponding element of vy, vector(real(8)), value :: vx, vy vector(real(8)), value :: vx, vy computes sqrt(vx*vx +vy*vy). lgammad4 lgammaf4 Computes the natural logarithm vector(real(8)) function lgammad4(vx) vector(real(8)) function lgammaf4(vx) of the absolute value of the vector(real(8)), value :: vx vector(real(8)), value :: vx Gamma function of each element of vx . logd4 logf4 Computes the natural logarithm vector(real(8)) function logd4(vx) vector(real(8)) function logf4(vx) of each element of vx. vector(real(8)), value :: vx vector(real(8)), value :: vx log2d4 log2f4 Computes the base-2 logarithm vector(real(8)) function log2d4(vx) vector(real(8)) function log2f4(vx) of each element of vx. vector(real(8)), value :: vx vector(real(8)), value :: vx log10d4 log10f4 Computes the base-10 logarithm vector(real(8)) function log10d4(vx) vector(real(8)) function log10f4(vx) of each element of vx. vector(real(8)), value :: vx vector(real(8)), value :: vx log1pd4 log1pf4 Computes the natural logarithm vector(real(8)) function log1pd4(vx) vector(real(8)) function log1pf4(vx) of each element of (vx +1). vector(real(8)), value :: vx vector(real(8)), value :: vx log21pd4 log21pf4 Computes the base-2 logarithm vector(real(8)) function log21pd4(vx) vector(real(8)) function log21pf4(vx) of each element of (vx +1). vector(real(8)), value :: vx vector(real(8)), value :: vx powd4 powf4 Computes each element of vx vector(real(8)) function powd4(vx, vy) vector(real(8)) function powf4(vx, vy) raised to the power of the vector(real(8)), value :: vx, vy vector(real(8)), value :: vx, vy corresponding element of vy. qdrtd4 qdrtf4 Computes the quad root of each vector(real(8)) function qdrtd4(vx) vector(real(8)) function qdrtf4(vx) element of vx. vector(real(8)), value :: vx vector(real(8)), value :: vx rcbrtd4 rcbrtf4 Computes the reciprocal of the vector(real(8)) function rcbrtd4(vx) vector(real(8)) function rcbrtf4(vx) cube root of each element of vx. vector(real(8)), value :: vx vector(real(8)), value :: vx recipd4 recipf4 Computes the reciprocal of each vector(real(8)) function recipd4(vx) vector(real(8)) function recipf4(vx) element of vx. vector(real(8)), value :: vx vector(real(8)), value :: vx recip_fastd4 (3) recip_fastf4 (4) Computes the reciprocal of each vector(real(8)) function recip_fastd4(vx) vector(real(8)) function recip_fastf4(vx) element of vx. vector(real(8)), value :: vx vector(real(8)), value :: vx rqdrtd4 rqdrtf4 Computes the reciprocal of the vector(real(8)) function rqdrtd4(vx) vector(real(8)) function rqdrtf4(vx) quad root of each element of vx. vector(real(8)), value :: vx vector(real(8)), value :: vx rsqrtd4 rsqrtf4 Computes the reciprocal of the vector(real(8)) function rsqrtd4(vx) vector(real(8)) function rsqrtf4(vx) square root of each element of vector(real(8)), value :: vx vector(real(8)), value :: vx vx. sincosd4 sincosf4 Computes the sine and cosine of subroutine sincosd4(vx, vs, vc) subroutine sincosf4(vx, vs, vc) each element of vx. vector(real(8)), value :: vx vector(real(8)), value :: vx vector(real(8)) vs, vc vector(real(8)) vs, vc sind4 sinf4 Computes the sine of each vector(real(8)) function sind4(vx) vector(real(8)) function sinf4(vx) element of vx. vector(real(8)), value :: vx vector(real(8)), value :: vx sinhd4 sinhf4 Computes the hyperbolic sine of vector(real(8)) function sinhd4(vx) vector(real(8)) function sinhf4(vx) each element of vx.i vector(real(8)), value :: vx vector(real(8)), value :: vx sqrtd4 sqrtf4 Computes the square root of vector(real(8)) function sqrtd4(vx) vector(real(8)) function sqrtf4(vx) each element of vx. vector(real(8)), value :: vx vector(real(8)), value :: vx tand4 tanf4 Computes the tangent of each vector(real(8)) function tand4(vx) vector(real(8)) function tanf4(vx) element of vx. vector(real(8)), value :: vx vector(real(8)), value :: vx tanhd4 tanhf4 Computes the hyperbolic vector(real(8)) function tanhd4(vx) vector(real(8)) function tanhf4(vx) tangent of each element of vx. vector(real(8)), value :: vx vector(real(8)), value :: vx
Notes: 1. div_fastd4 arguments must satisfy all the following conditions for each element xi of x and yi of y: v 2-1021 ≤|yi|≤21020 v If xi is not zero, 2-969 ≤|xi|<∞and 2-1020 ≤ |xi/yi| ≤ 21022 2. div_fastf4 arguments must satisfy all the following conditions for each element xi of x and yi of y: v 2-125 ≤|yi|≤2124
Chapter 6. High performance libraries 63 v If xi is not zero, 2-102 ≤|xi|<∞and 2-124 ≤ |xi/yi| ≤ 2126 3. recip_fastd4 arguments must satisfy the following condition for each element xi of x: v 2-1021 ≤ |xi| ≤ 21020 4. recip_fastf4 arguments must satisfy the following condition for each element xi of x: v 2-125 ≤ |xi| ≤ 2124 Compiling and linking a program with MASS To compile an application that calls the functions in the MASS libraries, specify one or more of the following keywords on the -l linker option: v mass v massv v mass_simd
For example, if the MASS libraries are installed in the default directory, you can specify one of the following: Link with scalar library libmass.a and vector library libmassv.a bgxlf progf.f -o progf -lmass -lmassv Link with SIMD library libmass_simd.a bgxlf progf.f -o progf -lmass_simd Using libmass.a with the math system library If you want to use the libmass.a scalar library for some functions and the normal math library libm.a for other functions, follow this procedure to compile and link your program: 1. Use the ar command to extract the object files of the desired functions from libmass.a. For most functions, the object file name is the function name followed by .s64.o. 1 For example, to extract the object file for the tan function, the command would be: ar -x tan.s64.o libmass.a 2. Archive the extracted object files into another library: ar -qv libfasttan.a tan.s64.o ranlib libfasttan.a 3. Create the final executable using bgxlf, specifying -lfasttan instead of -lmass: bgxlf sample.f -o sample -Ldir_containing_libfasttan -lfasttan
This links only the tan function from MASS (now in libfasttan.a) and the remainder of the math functions from the standard system library.
Exceptions: 1. The sin and cos functions are both contained in the object file sincos.s64.o. The cosisin and sincos functions are both contained in the object file cosisin.s64.o. 2. The XL Fortran ** (exponentiation) operator is contained in the object file dxy.s64.o.
Note: The cos and sin functions will both be exported if either one is exported. cosisin and sincos will both be exported if either one is exported.
64 XL Fortran: Optimization and Programming Guide Using the Basic Linear Algebra Subprograms – BLAS Four Basic Linear Algebra Subprograms (BLAS) functions are shipped with XL Fortran in the libxlopt library.
The functions consist of the following: v SGEMV (single-precision) and DGEMV (double-precision), which compute the matrix-vector product for a general matrix or its transpose v SGEMM (single-precision) and DGEMM (double-precision), which perform combined matrix multiplication and addition for general matrices or their transposes
Note: Some error-handling code has been removed from the BLAS functions in libxlopt, and no error messages are emitted for calls to the these functions.
“BLAS function syntax” describes the interfaces for the XL Fortran BLAS functions, which are similar to those of the equivalent BLAS functions shipped in IBM's Engineering and Scientific Subroutine Library (ESSL); for more detailed information and examples of usage of these functions, you may wish to consult the Engineering and Scientific Subroutine Library Guide and Reference, available at the Engineering and Scientific Subroutine Library (ESSL) and Parallel ESSL web page.
“Linking the libxlopt library” on page 67 describes how to link to the XL Fortran libxlopt library if you are also using a third-party BLAS library. BLAS function syntax The interfaces for the SGEMV and DGEMV functions are as follows: CALL SGEMV(trans, m, n, alpha, a, lda, x, incx, beta, y, incy) CALL DGEMV(trans, m, n, alpha, a, lda, x, incx, beta, y, incy)
The parameters are as follows: trans is a single character indicating the form of the input matrix a, where: v ’N’ or ’n’ indicates that a is to be used in the computation v ’T’ or ’t’ indicates that the transpose of a is to be used in the computation m represents: v the number of rows in input matrix a v the length of vector y,if’N’ or ’n’ is used for the trans parameter v the length of vector x,if’T’ or ’t’ is used for the trans parameter The number of rows must be greater than or equal to zero, and less than or equal to the leading dimension of the matrix a (specified in lda) n represents: v the number of columns in input matrix a v the length of vector x,if’N’ or ’n’ is used for the trans parameter v the length of vector y,if’T’ or ’t’ is used for the trans parameter The number of columns must be greater than or equal to zero. alpha is the scaling constant α
Chapter 6. High performance libraries 65 a is the input matrix of single-precision (for SGEMV) or double-precision (for DGEMV) real values lda is the leading dimension of the array specified by a. The number of rows must be greater than or equal to zero, and less than the leading dimension of the matrix a (specified in lda). x is the input vector of single-precision (for SGEMV) or double-precision (for DGEMV) real values. incx is the stride for vector x. It can have any value. beta is the scaling constant β y is the output vector of single-precision (for SGEMV) or double-precision (for DGEMV) real values. incy is the stride for vector y. It must not be zero.
Note: Vector y must have no common elements with matrix a or vector x; otherwise, the results are unpredictable.
The prototypes for the SGEMM and DGEMM functions are as follows: CALL SGEMM(transa, transb, l, n, m, alpha, a, lda, b, ldb, beta, c, ldc) CALL DGEMM(transa, transb, l, n, m, alpha, a, lda, b, ldb, beta, c, ldc)
The parameters are as follows: transa is a single character indicating the form of the input matrix a, where: v ’N’ or ’n’ indicates that a is to be used in the computation v ’T’ or ’t’ indicates that the transpose of a is to be used in the computation transb is a single character indicating the form of the input matrix b, where: v ’N’ or ’n’ indicates that b is to be used in the computation v ’T’ or ’t’ indicates that the transpose of b is to be used in the computation l represents the number of rows in output matrix c. The number of rows must be less than or equal to the leading dimension of c. n represents the number of columns in output matrix c. The number of columns must be greater than or equal to zero. m represents: v the number of columns in matrix a,if’N’ or ’n’ is used for the transa parameter v the number of rows in matrix a,if’T’ or ’t’ is used for the transa parameter and: v the number of rows in matrix b,if’N’ or ’n’ is used for the transb parameter v the number of columns in matrix b,if’T’ or ’t’ is used for the transb parameter m must be greater than or equal to zero.
66 XL Fortran: Optimization and Programming Guide alpha is the scaling constant α a is the input matrix a of single-precision (for SGEMM) or double-precision (for DGEMM) real values lda is the leading dimension of the array specified by a. The leading dimension must be greater than zero. If transa is specified as ’N’ or ’n’, the leading dimension must be greater than or equal to 1. If transa is specified as ’T’ or ’t’, the leading dimension must be greater than or equal to the value specified in m. b is the input matrix b of single-precision (for SGEMM) or double-precision (for DGEMM) real values. ldb is the leading dimension of the array specified by b. The leading dimension must be greater than zero. If transb is specified as ’N’ or ’n’, the leading dimension must be greater than or equal to the value specified in m.Iftransa is specified as ’T’ or ’t’, the leading dimension must be greater than or equal to the value specified in n. beta is the scaling constant β c is the output matrix c of single-precision (for SGEMM) or double-precision (for DGEMM) real values. ldc is the leading dimension of the array specified by c. The leading dimension must be greater than zero. If transb is specified as ’N’ or ’n’, the leading dimension must be greater than or equal to the value specified in l.
Note: Matrix c must have no common elements with matrices a or b; otherwise, the results are unpredictable. Linking the libxlopt library By default, the libxlopt library is linked with any application you compile with XL Fortran. However, if you are using a third-party BLAS library, but want to use the BLAS routines shipped with libxlopt, you must specify the libxlopt library before any other BLAS library on the command line at link time. For example, if your other BLAS library is called libblas, you would compile your code with the following command: bgxlf app.f -lxlopt -lblas
The compiler will call the SGEMV, DGEMV, SGEMM, and DGEMM functions from the libxlopt library, and all other BLAS functions in the libblas library.
Chapter 6. High performance libraries 67 68 XL Fortran: Optimization and Programming Guide Chapter 7. Parallel programming with XL Fortran
Parallel programming with XL Fortran involves a combination of compiling, setting of runtime options, and optimization of your code, by incorporating SMP directives and by using the pthreads library module.
XL Fortran supports the OpenMP specification, as understood and interpreted by IBM as well as the POSIX 1003.1-1996 standard, the transactional memory feature, and the thread-level speculative execution feature.
Note: IBM implementation of OpenMP in XL Fortran is the extension to the standard Fortran language.
Thread-level speculative execution Thread-level speculative execution overcomes the analysis problems of compiler-directed code parallelization.
In compiler-directed code parallelization, the compiler must analyze the code to ensure that the code can be parallelized. However, the C-style pointers and array subscript arithmetic operations hamper the compiler analysis. Furthermore, the compiler generates parallel code only if dependencies can be ruled out; otherwise, it generates sequential code.
The XL Fortran compiler supports the programming paradigm of thread-level speculative execution. Thread-level speculative execution uses hardware support that dynamically detects thread conflicts and rolls back conflicting threads for re-execution. You can get significant performance gains in your applications by adding the compiler directives of thread-level speculative execution to the existing program code.
Thread-level speculative execution is enabled with the "-qsmp=speculative" compiler option. Rules for committing data
With thread-level speculative execution, tasks are committed according to the following rules: v Before a task is committed, the data is in a speculative state. v Tasks are committed in program order. v Therefore, a later task in program order can only be committed when all the earlier tasks have been committed. If a thread running a task encounters a conflict, all the threads running later tasks must roll back and retry. Eventually, all tasks are committed. Thread-level speculative execution and OpenMP
All the OpenMP restrictions that apply to the PARALLEL DO and PARALLEL SECTIONS directives apply to the SPECULATIVE DO and SPECULATIVE SECTIONS directives.
© Copyright IBM Corp. 1990, 2012 69 Speculative threads are not able to detect access conflicts in OpenMP THREADPRIVATE data. Accessing such data inside regions of thread-level speculative execution does not guarantee the same behavior as regions being run by one thread.
For the restrictions of the OpenMP loop construct, see the Loop Construct section in the OpenMP specification 3.1.
For the restrictions of the OpenMP sections construct, see the sections Construct section in the OpenMP specification 3.1. Related information v The "-qsmp" compiler option v Routines for thread-level speculative execution v Environment variables for thread-level speculative execution v SPECULATIVE DO / END SPECULATIVE DO v SPECULATIVE SECTIONS v THREADPRIVATE
Transactional memory Transactional memory is a model for controlling concurrent memory accesses in the scope of parallel programming.
In parallel programming, concurrency control ensures that threads running in parallel do not update the same resources at the same time. Traditionally, the concurrency control of shared memory data is through locks, for example, mutex locks. A thread acquires a lock before modifying the shared data, and releases the lock afterward. A lock-based synchronization can lead to some performance issues because threads might need to wait to update lock-protected data.
Transactional memory is an alternative to lock-based synchronization. It attempts to simplify parallel programming by grouping read and write operations and running them like a single operation. Transactional memory is like database transactions where all shared memory accesses and their effects are either committed all together or discarded as a group. All threads can enter the critical region simultaneously. If there are conflicts in accessing the shared memory data, threads try accessing the shared memory data again or are stopped without updating the shared memory data. Therefore, transactional memory is also called a lock-free synchronization. Transactional memory can be a competitive alternative to lock-based synchronization.
A transactional memory system must hold the following properties across the entire execution of a concurrent program: Atomicity All speculative memory updates of a transaction are either committed or discarded as a unit. Consistency The memory operations of a transaction take place in order. Transactions are committed one transaction at a time. Isolation Memory updates are not visible outside of a transaction until the transaction commits data.
70 XL Fortran: Optimization and Programming Guide Transactional memory on Blue Gene/Q
On Blue Gene/Q, the transactional memory model is implemented in the hardware to access all the memory up to the 16 GB boundary.
Transactions are implemented through regions of code that you can designate to be single operations for the system. The regions of code that implement the transactions are called transactional atomic regions.
Transactional memory is enabled with the "-qtm" compiler option, and requires thread safe compilation mode. Execution modes
When transactional memory is activated on Blue Gene/Q, transactions are run in one of the following operating modes: v Speculation mode – Long running speculation mode (default) – Short running speculation mode v Irrevocable mode Each mode applies to an entire transactional atomic region. Speculation mode Under speculation mode, Kernel address space, devices I/Os, and most memory-mapped I/Os are protected from the irrevocable actions except when the safe_mode clause is specified. The transaction goes into irrevocable mode if such an action occurs to guarantee the correct result. Blue Gene/Q supports two hardware implementation of transaction memory: long and short running speculation mode. If the transactional atomic region is large and many reuses are among the references inside the transaction, it is recommended that you use the default long running speculation mode. Otherwise, use the short running speculation mode. Irrevocable mode System calls, irrevocable operations such as I/O operations, and OpenMP constructs trigger transactions to go into irrevocable mode, which serializes transactions. Transactions are also running in irrevocable mode when the maximum number of transaction rollbacks has been reached. Under irrevocable mode, each memory update of a thread is committed instantaneously instead of at the end of the transaction. Therefore, memory updates are immediately visible to other threads. If the transaction becomes irrevocable, the threads run nonspeculatively. Using variables and synchronization constructs with transactional memory
The semantics of transactional memory ensure that the effects of transactions of a thread are visible to other threads only after the transactions commit data or become irrevocable. When you use variables or synchronization constructs inside transactions, be careful when the volatile and regular variables that are visible to other threads are updated.
Chapter 7. Parallel programming with XL Fortran 71 Data races when using transactional memory
A data race might happen if a memory location is accessed concurrently from both the following types of code sections: v A transactional atomic region that is not nested in other critical sections v A lock-based critical section of another thread
For example, the atomicity of a lock-based critical section might be broken when the transaction happens in the middle of the critical section. The atomicity of the transaction might also be broken if the transaction becomes irrevocable and is interleaved with the critical section.
The data race happens because each transactional atomic region can be thought of as using a different lock. In contrast, the !$omp critical directive uses one lock for all critical regions in the same parallel region. Related information v The "-qtm" compiler option v Routines for transactional memory v Environment variables for transactional memory v TM_ATOMIC / END TM_ATOMIC
Profiler for OpenMP Profiler for OpenMP (POMP) is a profiling mechanism for OpenMP runtime. It inserts callbacks at key points in an OpenMP program; for example, when a parallel region is entered or exited. In these callbacks, you can run your code for various purposes, such as to increment profiling counters, or record timestamps.
For further details about POMP, see the Profiler for OpenMP section in XL C/C++ Optimization and Programming Guide.
Compiling your parallelized code To compile parallelized code, you must specify the -qsmp compiler option. When compiling with -qsmp, the driver links the libraries found on the smplibraries line in the active stanza of your configuration file.
If you specify -qsmp, you must use an appropriate invocation command. Use any of the following invocations to compile SMP code or to ensure that the compiler links threadsafe libraries: v bgxlf_r v bgxlf90_r v bgxlf95_r v bgxlf2003_r v bgxlf2008_r
72 XL Fortran: Optimization and Programming Guide Related reference:
See -qsmp in the Compiler Reference The _OPENMP C preprocessor macro and conditional compilation You can use sentinels to mark specific lines of an XL Fortran program for conditional compilation. This allows you to port code that contains statements that are only valid or applicable in an SMP environment to a non-SMP environment. You can do this using conditional compilation lines, or the _OPENMP C preprocessor macro. This macro is defined when the C preprocessor is invoked and you specify the -qsmp=omp compiler option. See Passing Fortran files through the C preprocessor in the Editing, Compiling, Linking, and Running XL Fortran Programs section of the XL Fortran Compiler Reference for an example of using this macro.
The following example uses conditional compilation lines to hide OpenMP runtime routines. You cannot easily compile code that calls OpenMP runtime routines in a non-OpenMP environment without using conditional compilation. Since calls to the runtime routines are not directives, they cannot be hidden by the !$OMP trigger. If you do not compile the example with -qsmp=omp, the variable that stores the number of threads is assigned the value of 8.
Example of conditional compilation lines PROGRAM PAR_MAT_MUL !$ USE OMP_LIB IMPLICIT NONE INTEGER(KIND=8) ::I,J,NTHREADS INTEGER(KIND=8),PARAMETER ::N=60 INTEGER(KIND=8),DIMENSION(N,N) ::AI,BI,CI INTEGER(KIND=8) ::SUMI
COMMON/DATA/ AI,BI,CI !$OMP THREADPRIVATE (/DATA/)
!$OMP PARALLEL FORALL(I=1:N,J=1:N) AI(I,J) = (I-N/2)**2+(J+N/2) FORALL(I=1:N,J=1:N) BI(I,J) = 3-((I/2)+(J-N/2)**2) !$OMP MASTER NTHREADS=8 !$ NTHREADS=OMP_GET_NUM_THREADS() !$OMP END MASTER !$OMP END PARALLEL
!$OMP PARALLEL DEFAULT(PRIVATE),COPYIN(AI,BI),SHARED(NTHREADS) !$OMP DO DO I=1,NTHREADS CALL IMAT_MUL(SUMI) ENDDO !$OMP END DO !$OMP END PARALLEL
END
For information on using sentinels, see Conditional compilation in the XL Fortran Language Reference.
Chapter 7. Parallel programming with XL Fortran 73 Setting runtime options When you write parallel code, set the necessary XLSMPOPTS environment variables, and the environment variables for OpenMP, transactional memory, and thread-level speculative execution. XLSMPOPTS The XLSMPOPTS environment variable allows you to specify options that affect SMP execution. You can declare XLSMPOPTS by using the following bash command format:
:
XLSMPOPTS= runtime_option_name = option_setting ""
You can specify option names and settings in uppercase or lowercase. You can add blanks before and after the colons and equal signs to improve readability. However, if the XLSMPOPTS option string contains imbedded blanks, you must enclose the entire option string in double quotation marks (").
You can specify the following runtime options with the XLSMPOPTS environment variable: schedule Selects the scheduling type and chunk size to be used as the default at run time. The scheduling type that you specify will only be used for loops that were not already marked with a scheduling type at compilation time. Work is assigned to threads in a different manner, depending on the scheduling type and chunk size used. A brief description of the scheduling types and their influence on how work is assigned follows: dynamic or guided The runtime library dynamically schedules parallel work for threads on a "first-come, first-do" basis. "Chunks" of the remaining work are assigned to available threads until all work has been assigned. Work is not assigned to threads that are asleep. static Chunks of work are assigned to the threads in a "round-robin" fashion. Work is assigned to all threads, both active and asleep. The system must activate sleeping threads in order for them to complete their assigned work. affinity The runtime library performs an initial division of the iterations into number_of_threads partitions. The number of iterations that these partitions contain is: CEILING(number_of_iterations / number_of_threads)
These partitions are then assigned to each of the threads. It is these partitions that are then subdivided into chunks of iterations. If a thread is asleep, the threads that are active will complete their assigned partition of work. Choosing chunking granularity is a tradeoff between overhead and load balancing. The syntax for this option is schedule=suboption, where the suboptions are defined as follows:
74 XL Fortran: Optimization and Programming Guide affinity[=n] As described previously, the iterations of a loop are initially divided into partitions, which are then preassigned to the threads. Each of these partitions is then further subdivided into chunks that contain n iterations. If you have not specified n, a chunk consists of CEILING(number_of_iterations_left_in_local_partition / 2) loop iterations. When a thread becomes available, it takes the next chunk from its preassigned partition. If there are no more chunks in that partition, the thread takes the next available chunk from a partition preassigned to another thread. auto With auto, scheduling is delegated to the compiler and runtime system. The compiler and runtime system can choose any possible mapping of iterations to threads (including all possible valid schedules) and these may be different in different loops. Do not specify chunk size (n) when you use auto. If chunk size (n) is specified, the compiler issues a severe error message.
Note: When both the -qsmp=schedule option and OMP_SCHEDULE are used, the option will override the environment variable. dynamic[=n] The iterations of a loop are divided into chunks that contain n iterations each. If you have not specified n, a chunk consists of CEILING(number_of_iterations / number_of_threads) iterations. guided[=n] The iterations of a loop are divided into progressively smaller chunks until a minimum chunk size of n loop iterations is reached. If you have not specified n, the default value for n is 1 iteration. The first chunk contains CEILING(number_of_iterations / number_of_threads) iterations. Subsequent chunks consist of CEILING(number_of_iterations_left / number_of_threads) iterations. static[=n] The iterations of a loop are divided into chunks that contain n iterations. Threads are assigned chunks in a "round-robin" fashion. This is known as block cyclic scheduling. If the value of n is 1, the scheduling type is specifically referred to as cyclic scheduling. If you have not specified n, the chunks will contain CEILING(number_of_iterations / number_of_threads) iterations. Each thread is assigned one of these chunks. This is known as block scheduling. If you have not specified schedule, the default is set to schedule=static, resulting in block scheduling. For more information, see the description of the SCHEDULE directive in the XL Fortran Language Reference. Parallel execution options parthds=num Specifies the number of threads (num) to be used for parallel execution of code that you compiled with the -qsmp option. By default, this is equal to the number of online processors. There are some applications that cannot use more than some maximum number of processors. There are also some applications that can achieve performance gains if they use more threads than there are processors.
Chapter 7. Parallel programming with XL Fortran 75 This option allows you full control over the number of execution threads. The default value for num is 1 if you did not specify -qsmp. Otherwise, it is the number of online processors on the machine. For more information, see the NUM_PARTHDS intrinsic function in the XL Fortran Language Reference. usrthds=num Specifies the maximum number of threads (num) that you expect your code will explicitly create if the code does explicit thread creation. The default value for num is 0. For more information, see the NUM_PARTHDS intrinsic function in the XL Fortran Language Reference. stack=num Specifies the largest amount of space in bytes (num) that a thread's stack will need. The default value for num is 4194304. Set stack=num so it is within the acceptable upper limit. num can be up to the limit imposed by system resources or the stack size ulimit, whichever is smaller. An application that exceeds the upper limit may cause a segmentation fault. stackcheck[=num] Enables stack overflow checking for worker threads at runtime. num is the size in bytes that you specify; when the remaining stack size is less than num, a runtime warning message is issued. If you do not specify a value for num, the default value is 4096 bytes. Note that this option only has an effect when -qsmp=stackcheck has also been specified at compile time. See -qsmp in the XL Fortran Compiler Reference for more information. startproc=cpu_id Enables thread binding and specifies the cpu_id to which the first thread binds. If the value provided is outside the range of available processors, the SMP run time issues a warning message and no threads are bound. procs=cpu_id[,cpu_id,...] Enables thread binding and specifies a list of cpu_id to which the threads are bound. If the number of CPU IDs specified is less than the number of threads used by the program, the remaining threads are not bound. stride=num Specifies the increment used to determine the cpu_id to which subsequent threads bind. num must be greater than or equal to 1. If the value provided causes a thread to bind to a CPU outside the range of available processors, a warning message is issued and no threads are bound. Performance tuning options When a thread completes its work and there is no new work to do, it can go into either a "busy-wait" state or a "sleep" state. In "busy-wait", the thread keeps executing in a tight loop looking for additional new work. This state is highly responsive but harms the overall utilization of the system. When a thread sleeps, it completely suspends execution until another thread signals it that there is work to do. This state provides better utilization of the system but introduces extra overhead for the application.
76 XL Fortran: Optimization and Programming Guide The xlsmp runtime library routines use both "busy-wait" and "sleep" states in their approach to waiting for work. You can control these states with the spins, yields, and delays options. During the busy-wait search for work, the thread repeatedly scans the work queue up to num times, where num is the value that you specified for the option spins. If a thread cannot find work during a given scan, it intentionally wastes cycles in a delay loop that executes num times, where num is the value that you specified for the option delays. This delay loop consists of a single meaningless iteration. The length of actual time this takes will vary among processors. If the value spins is exceeded and the thread still cannot find work, the thread will yield the current time slice (time allocated by the processor to that thread) to the other threads. The thread will yield its time slice up to num times, where num is the number that you specified for the option yields. If this value num is exceeded, the thread will go to sleep. In summary, the ordered approach to looking for work consists of the following steps: 1. Scan the work queue for up to spins number of times. If no work is found in a scan, then loop delays number of times before starting a new scan. 2. If work has not been found, then yield the current time slice. 3. Repeat the above steps up to yields number of times. 4. If work has still not been found, then go to sleep. The syntax for specifying these options is as follows: spins[=num] where num is the number of spins before a yield. The default value for spins is 100. yields[=num] where num is the number of yields before a sleep. The default value for yields is 10. delays[=num] where num is the number of delays while busy-waiting. The default value for delays is 500. Zero is a special value for spins and yields, as it can be used to force complete busy-waiting. Normally, in a benchmark test on a dedicated system, you would set both options to zero. However, you can set them individually to achieve other effects. For instance, on a dedicated 8-way SMP, setting these options to the following: parthds=8 : schedule=dynamic=10 : spins=0 : yields=0
results in one thread per CPU, with each thread assigned chunks consisting of 10 iterations each, with busy-waiting when there is no immediate work to do. Options to enable and control dynamic profiling You can use dynamic profiling to reevaluate the compiler's decision to parallelize loops in a program. The three options you can use to do this are: parthreshold, seqthreshold, and profilefreq. parthreshold=num Specifies the time, in milliseconds, below which each loop must execute serially. If you set parthreshold to 0, every loop that has been parallelized by the compiler will execute in parallel. The default setting is 0.2 milliseconds, meaning that if a loop requires fewer than 0.2 milliseconds to execute in parallel, it should be serialized.
Chapter 7. Parallel programming with XL Fortran 77 Typically, parthreshold is set to be equal to the parallelization overhead. If the computation in a parallelized loop is very small and the time taken to execute these loops is spent primarily in the setting up of parallelization, these loops should be executed sequentially for better performance. seqthreshold=num Specifies the time, in milliseconds, beyond which a loop that was previously serialized by the dynamic profiler should revert to being a parallel loop. The default setting is 5 milliseconds, meaning that if a loop requires more than 5 milliseconds to execute serially, it should be parallelized. seqthreshold acts as the reverse of parthreshold. profilefreq=num Specifies the frequency with which a loop should be revisited by the dynamic profiler to determine its appropriateness for parallel or serial execution. Loops in a program can be data dependent. The loop that was chosen to execute serially with a pass of dynamic profiling may benefit from parallelization in subsequent executions of the loop, due to different data input. Therefore, you need to examine these loops periodically to reevaluate the decision to serialize a parallel loop at run time. The allowed values for this option are the numbers from 0 to 32. If you set profilefreq to one of these values, the following results will occur. v If profilefreq is 0, all profiling is turned off, regardless of other settings. The overheads that occur because of profiling will not be present. v If profilefreq is 1, loops parallelized automatically by the compiler will be monitored every time they are executed. v If profilefreq is 2, loops parallelized automatically by the compiler will be monitored every other time they are executed. v If profilefreq is greater than or equal to 2 but less than or equal to 32, each loop will be monitored once every nth time it is executed. v If profilefreq is greater than 32, then 32 is assumed. It is important to note that dynamic profiling is not applicable to user-specified parallel loops (for example, loops for which you specified the PARALLEL DO directive). BG_SMP_FAST_WAKEUP
The BG_SMP_FAST_WAKEUP environment variable enables or disables the fast wake-up support, which allows the SMP runtime to use the hardware mechanism that puts waiting threads to sleep and wakes them up efficiently.
NO BG_SMP_FAST_WAKEUP = YES
The default value is NO. The value is not case sensitive.
BG_SMP_FAST_WAKEUP=YES enables the fast wake-up support.
78 XL Fortran: Optimization and Programming Guide When a thread is waiting for work or spinning at a barrier, it puts itself to sleep so that it does not consume processing resources. Due to the Blue Gene/Q threading model, the sleeping cannot be interrupted. This causes an issue when the hardware threads are oversubscribed; that is, multiple user threads are bound to the same hardware thread. This issue might cause deadlocks between user threads, in particular between OpenMP threads in the SMP runtime.
To avoid oversubscription of hardware threads when the fast wake-up support is enabled, you must also set the OMP_PROC_BIND environment variable to TRUE. This setting ensures that each user thread created by the SMP runtime is bound to a different hardware thread. Related information v OMP_PROC_BIND BG_SPEC_SCRUB_CYCLE
The BG_SPEC_SCRUB_CYCLE environment variable sets the L2 background scrub cycle, which determines how quick a speculation ID can be reclaimed by the L2 and used again for speculation.
66 BG_SPEC_SCRUB_CYCLE = n
The default value is 66. The valid value range is 6-100.
Setting the scrub cycle low means the L2 needs to perform scrub more frequently.
The interval defines the period in 800MHz cycles (two processor cycles) for directory look-ups of the scrub process. For the default setting, the L2 examines one of the 1024 sets every 132 processor cycles, resulting in a complete cache scrub every 135168 processor cycles.
After a speculation ID has been used and committed or recycled, it takes at most 135168 cycles until the ID is available again for allocation. It can take less if all sets that were accessed with the ID are visited by either core memory accesses or the scrub earlier.
For example, programming the interval to 6 changes the duration for a full cache scrub from 135168 to 12288 processor cycles. Environment variables for OpenMP The following environment variables, which are included in the OpenMP standard, allow you to control the execution of parallel code.
Note: If you specify both the XLSMPOPTS environment variable and an OpenMP environment variable, the OpenMP environment variable takes precedence. OMP_DYNAMIC
The OMP_DYNAMIC environment variable enables or disables dynamic adjustment of the number of threads available for the execution of parallel regions. The syntax is as follows:
Chapter 7. Parallel programming with XL Fortran 79 OMP_DYNAMIC= TRUE FALSE
If you set this environment variable to TRUE, the runtime environment can adjust the number of threads it uses for executing parallel regions so that it makes the most efficient use of system resources. If you set this environment variable to FALSE, dynamic adjustment is disabled.
The default value for OMP_DYNAMIC is FALSE. If your code needs to use a specific number of threads to run correctly, you should disable dynamic thread adjustment.
The omp_set_dynamic subroutine takes precedence over the OMP_DYNAMIC environment variable. OMP_MAX_ACTIVE_LEVELS
The OMP_MAX_ACTIVE_LEVELS environment variable controls the maximum number of nested active parallel regions. The syntax is as follows:
OMP_MAX_ACTIVE_LEVELS= n
n is the maximum number of nested active parallel regions. It must be a positive scalar integer. XL Fortran does not support OpenMP nested parallelism. This environment variable has no effects to the nested parallel constructs in the program. OMP_NESTED The OMP_NESTED environment variable enables or disables nested parallelism. The syntax is as follows:
OMP_NESTED= TRUE FALSE
If you set this environment variable to TRUE, nested parallelism is enabled. This means that the runtime environment might deploy extra threads to form the team of threads for the nested parallel region. If you set this environment variable to FALSE, nested parallelism is disabled.
The default value for OMP_NESTED is FALSE.
The omp_set_nested subroutine takes precedence over the OMP_NESTED environment variable.
Currently, XL Fortran does not support OpenMP nested parallelism. OMP_NUM_THREADS The OMP_NUM_THREADS environment variable sets the number of threads to use for parallel regions. The syntax of the environment variable is as follows:
OMP_NUM_THREADS= num_list
80 XL Fortran: Optimization and Programming Guide num_list A list of one or more positive integer values separated by commas.
If you do not set the OMP_NUM_THREADS environment variable, the number of processors available is the default value to form a new team for the first encountered parallel construct. By default, any nested constructs are run by one thread.
If num_list contains a single value, dynamic adjustment of the number of threads is enabled (OMP_DYNAMIC is set to true), a parallel construct without a NUM_THREADS clause is encountered, the value is the maximum number of threads that can be used to form a new team for the encountered parallel construct.
If num_list contains a single value, dynamic adjustment of the number of threads is not enabled (OMP_DYNAMIC is set to false), a parallel construct without a NUM_THREADS clause is encountered, the value is the exact number of threads that can be used to form a new team for the encountered parallel construct.
If num_list contains multiple values, dynamic adjustment of the number of threads is enabled (OMP_DYNAMIC is set to true), a parallel construct without a NUM_THREADS clause is encountered, the first value is the maximum number of threads that can be used to form a new team for the encountered parallel construct. After the encountered construct is entered, the first value is removed and the remaining values form a new num_list. The new num_list is in turn used in the same way for any closely nested parallel constructs inside the encountered parallel construct.
If num_list contains multiple values, dynamic adjustment of the number of threads is not enabled (OMP_DYNAMIC is set to false), a parallel construct without a NUM_THREADS clause is encountered, the first value is the exact number of threads that can be used to form a new team for the encountered parallel construct. After the encountered construct is entered, the first value is removed and the remaining values form a new num_list. The new num_list is in turn used in the same way for any closely nested parallel constructs inside the encountered parallel construct.
Note: If the number of parallel regions is equal to or greater than the number of values in num_list, the omp_get_max_threads routine returns the last value of num_list in the parallel region.
If the number of threads requested exceeds the system resources available, the program stops.
The omp_set_num_threads routine sets the first value of num_list. The omp_get_max_threads routine returns the first value of num_list.
If you specify the number of threads for a given parallel region more than once with different settings, the compiler uses the following precedence order to determine which setting takes effect: 1. The number of threads set using the NUM_THREADS clause takes precedence over that set using the omp_set_num_threads routine. 2. The number of threads set using the omp_set_num_threads routine takes precedence over that set using the OMP_NUM_THREADS environment variable.
Chapter 7. Parallel programming with XL Fortran 81 3. The number of threads set using the OMP_NUM_THREADS environment variable takes precedence over that set using the PARTHDS suboption of the XLSMPOPTS environment variable.
Note: In a given parallel region, the omp_get_max_threads routine returns the first value of num_list, even though the actual number of threads running that parallel region might be different from the first value of num_list.
The following example shows how you can set the OMP_NUM_THREADS environment variable. export OMP_NUM_THREADS=5,10 export OMP_DYNAMIC=false
! OMP_GET_MAX_THREADS() returns 5 threads !$omp parallel ! OMP_GET_MAX_THREADS() returns 10 threads !$omp parallel ! OMP_GET_MAX_THREADS() returns 10 threads !$omp parallel ! OMP_GET_MAX_THREADS() returns 10 threads !$omp end parallel !$omp end parallel !$omp end parallel OMP_PROC_BIND The OMP_PROC_BIND environment variable controls whether OpenMP threads can be moved between processors. The syntax of the environment variable is as follows:
OMP_PROC_BIND= TRUE FALSE
By default, the OMP_PROC_BIND environment variable is set to TRUE. If you set OMP_PROC_BIND to FALSE, the threads can be moved between processors.
Restriction: On the BG/Q system, the OMP_PROC_BIND environment variable is always set to TRUE. If you set OMP_PROC_BIND to FALSE, the compiler ignores it.
Note: The OMP_PROC_BIND environment variable provides a portable way to control whether OpenMP threads can be migrated. OMP_SCHEDULE The OMP_SCHEDULE environment variable applies to the PARALLEL DO and work-sharing DO directives that have a schedule type of RUNTIME. The syntax is as follows:
OMP_SCHEDULE= sched_type , chunk_size
sched_type is either AUTO, DYNAMIC, GUIDED,orSTATIC. See the “SCHEDULE” on page 167 clause for a description of these scheduling parameters. chunk_size is a positive, scalar integer that represents the chunk size.
82 XL Fortran: Optimization and Programming Guide This environment variable is ignored for the PARALLEL DO and work-sharing DO directives that have a schedule type other than RUNTIME.
If you do not specify a schedule type either at compile time through a directive, or at run time through the OMP_SCHEDULE environment variable or the SCHEDULE option of the XLSMPOPTS environment variable, the default schedule type is AUTO, which delegates scheduling decision to the compiler and runtime system. You cannot specify chunk_size when the schedule type is set to AUTO.
If you specify both the SCHEDULE option of the XLSMPOPTS environment variable and the OMP_SCHEDULE environment variable, the OMP_SCHEDULE environment variable takes precedence.
The following examples show how you can set the OMP_SCHEDULE environment variable: export OMP_SCHEDULE="DYNAMIC" export OMP_SCHEDULE="GUIDED,4" export OMP_SCHEDULE="STATIC" export OMP_SCHEDULE="AUTO" OMP_STACKSIZE
The OMP_STACKSIZE environment variable specifies the size of the stack for threads created by the OpenMP runtime. The syntax is as follows:
OMP_STACKSIZE= size sizeB sizeK sizeM sizeG size is a positive integer that specifies the size of the stack for threads that are created by the OpenMP runtime. "B", "K", "M", "G" are letters that specify whether the given size is in Bytes, Kilobytes, Megabytes, or Gigabytes.
If only size is specified and none of "B", "K", "M", "G" is specified, size is in Kilobytes by default. This environment variable does not control the size of the stack for the initial thread.
The value assigned to the OMP_STACKSIZE environment variable is case insensitive and might have leading and trailing white space. The following examples show how you can set the OMP_STACKSIZE environment variable. export OMP_STACKSIZE="10M" export OMP_STACKSIZE=" 10 M "
If the value of OMP_STACKSIZE is not set, the initial value is set to the default value for 64-bit mode, up to the limit imposed by system resources).
If the compiler cannot deliver the stack size specified by the environment variable, or if OMP_STACKSIZE does not conform to the valid format, the compiler sets the environment variable to the default value.
Chapter 7. Parallel programming with XL Fortran 83 The OMP_STACKSIZE environment variable takes precedence over the stack suboption of the XLSMPOPTS environment variable. OMP_THREAD_LIMIT
The OMP_THREAD_LIMIT environment variable sets the number of OpenMP threads to use for the whole program. The syntax is as follows:
OMP_THREAD_LIMIT= n
n The number of OpenMP threads to use for the whole program. It must be a positive scalar integer.
The value for OMP_THREAD_LIMIT is a positive integer.
If the OMP_THREAD_LIMIT environment variable is not set and the OMP_NUM_THREADS environment variable is set to a single value, the default value for OMP_THREAD_LIMIT is the value of OMP_NUM_THREADS or the number of available processors, whichever is greater.
If the OMP_THREAD_LIMIT environment variable is not set and the OMP_NUM_THREADS environment variable is set to a list, the default value for OMP_THREAD_LIMIT is the multiplication of all the numbers in the list or the number of available processors, whichever is greater.
If both the OMP_THREAD_LIMIT and OMP_NUM_THREADS environment variables are not set, the default value for OMP_THREAD_LIMIT is the number of available processors. OMP_WAIT_POLICY
The OMP_WAIT_POLICY environment variable provides hints about the preferred behavior of waiting threads during program execution. The syntax is as follows:
PASSIVE OMP_WAIT_POLICY= ACTIVE
Use ACTIVE if you want waiting threads to mostly be active. That is, the threads consume processor cycles while waiting. For example, waiting threads can spin while waiting. The ACTIVE wait policy is recommended for maximum performance on the dedicated machine.
Note: If the hardware threads are oversubscribed or multiple user threads are bound to the same hardware thread, OMP_WAIT_POLICY=ACTIVE might cause dead lock.
Use PASSIVE if you want waiting threads to mostly be passive. That is, the threads do not consume processor cycles while waiting. For example, waiting threads can sleep or yield the processor to other threads.
The default value of OMP_WAIT_POLICY is PASSIVE.
Note: If you set the OMP_WAIT_POLICY environment variable and specify the SPINS, YIELDS,orDELAYS suboptions of the XLSMPOPTS environment variable, OMP_WAIT_POLICY takes precedence.
84 XL Fortran: Optimization and Programming Guide Environment variables for thread-level speculative execution
The environment variables for thread-level speculative execution have no effect unless the thread-level speculative execution is enabled with the "-qsmp=speculative" compiler option. SE_MAX_NUM_ROLLBACK
The SE_MAX_NUM_ROLLBACK environment variable indicates the maximum number a speculative thread of a particular thread-level speculative execution region can roll back before the region is run by one thread.
10 SE_MAX_NUM_ROLLBACK = n
n The maximum number of times a thread for a speculative region can roll back before being run by a single thread. It is a positive integer and must not exceed 232. The default value is 10.
SE_MAX_NUM_ROLLBACK = 0 forces the region of thread-level speculative execution to be run by one thread.
Related information v Thread-level speculative execution SE_REPORT_STAT_ENABLE
The SE_REPORT_STAT_ENABLE environment variable enables or disables the statistics query routine for thread-level speculative execution.
NO SE_REPORT_STAT_ENABLE = YES
The default value is NO. The value is not case sensitive.
When SE_REPORT_STAT_ENABLE is set to YES, you can retrieve the statistics through the se_get_all_stats routine.
Related information v Routines for thread-level speculative execution SE_REPORT_NAME
The SE_REPORT_NAME environment variable specifies the name of the statistics log file and the location where it is created for thread-level speculative execution.
SE_REPORT_NAME = file_name
file_name The name of the log file, or the directory and name of the log file. If you specify the name only, the file is placed in the current working directory.
Chapter 7. Parallel programming with XL Fortran 85 If SE_REPORT_NAME is not set, the log file is placed in the current working directory and is named se_report.log.rank where rank is the MPI rank of the process that called the se_print_stats routine.
Related information v Routines for thread-level speculative execution SE_REPORT_LOG
The SE_REPORT_LOG environment variable specifies how to create the statistics log files for thread-level speculative execution.
SE_REPORT_LOG = SUMMARY FUNC ALL VERBOSE
The value is not case sensitive. SUMMARY The statistics log file is generated only at the end of the program. It contains the statistics of all the regions of thread-level speculative execution for all hardware threads. FUNC The statistics log file is generated and updated at each call to the se_print_stats routine. It contains all the statistics at the time when the se_print_stats routine is called. ALL The statistics log file is generated and updated at each call to the se_print_stats routine and at the end of the program. VERBOSE The statistics log file is generated and updated at each call to the se_print_stats routine and at the end of the program. The generated report file also includes the addresses of memory access conflicts during the speculation.
Related information v Routines for thread-level speculative execution Environment variables for transactional memory
The environment variables for transactional memory have no effect unless transactional memory is enabled with the "-qtm" compiler option. TM_MAX_NUM_ROLLBACK
The TM_MAX_NUM_ROLLBACK environment variable specifies the maximum number a thread of a particular transactional atomic region can roll back before the thread goes into irrevocable mode.
10 TM_MAX_NUM_ROLLBACK = n
n The maximum number of times a thread of a transactional atomic region
86 XL Fortran: Optimization and Programming Guide can roll back before executing in irrevocable mode. It is a positive integer and must not exceed 232. The default value is 10.
TM_MAX_NUM_ROLLBACK = 0 forces the thread to enter irrevocable mode. TM_ENABLE_INTERRUPT_ON_CONFLICT
The TM_ENABLE_INTERRUPT_ON_CONFLICT environment variable indicates whether to generate interrupts on conflicts between speculative access.
YES TM_ENABLE_INTERRUPT_ON_CONFLICT = NO
The default value is YES. The value is not case sensitive.
Conflict resolution usually happens at the end of transactions. Use this environment variable to generate interrupts for access conflicts in transactional atomic regions. This option is useful for long running transactions because the conflict resolution logic immediately starts inside the interrupt handler of the kernel.
Notes: v Interrupts are always generated for speculative buffer overflows, supervised-mode violations, or conflicts between speculative and non-speculative access. v Use with discretion if the TM_ENABLE_INTERRUPT_ON_CONFLICT environment variable is set to NO. If a transaction atomic region contains control flow that depends on some bad speculative data, the program may hang or the speculative thread may branch to invalid locations. TM_REPORT_STAT_ENABLE
The TM_REPORT_STAT_ENABLE environment variable enables or disables statistics query routines for transactional memory.
NO TM_REPORT_STAT_ENABLE = YES
The default value is NO. The value is not case sensitive.
When TM_REPORT_STAT_ENABLE is set to YES, you can retrieve the statistics through the following routines: v “tm_get_stats(stats)” on page 198 v “tm_get_all_stats(stats)” on page 199
Related information v Routines for transactional memory TM_REPORT_NAME
The TM_REPORT_NAME environment variable specifies the name of the statistics log file and the location where it is created for transactional memory.
Chapter 7. Parallel programming with XL Fortran 87 TM_REPORT_NAME = file_name
file_name The name of the log file, or the directory and name of the log file. If you specify the name only, the file is placed in the current working directory.
If TM_REPORT_NAME is not set, the log file is placed in the current working directory and is named tm_report.log.rank where rank in the extension is the MPI rank of the process that called the tm_print_stats.
Related information v Routines for transactional memory TM_REPORT_LOG
The TM_REPORT_LOG environment variable specifies how to create the statistics log file for transactional memory.
By default, no statistics log file is created.
TM_REPORT_LOG = SUMMARY FUNC ALL VERBOSE
The value is not case sensitive. SUMMARY The statistics log file is generated only at the end of the program. FUNC The statistics log file is generated and updated at each call to the tm_print_stats routine. ALL The statistics log file is generated and updated at each call to the tm_print_stats routine and at the end of the program. VERBOSE The statistics log file is generated and updated at each call to the tm_print_stats routine and at the end of the program. The generated report file contains the address of transactions that enter into irrevocable mode, and other useful information, such as the configuration of the runtime.
If TM_REPORT_LOG is set to either SUMMARY, ALL, or VERBOSE, a log file that contains the following statistics is also created when the program ends: v It contains the cumulative statistics for transactional atomic regions that each hardware thread has run. The statistics follow the Structure of statistic counters. v It contains a summary that shows the sum of all the statistic counters for all the hardware threads.
Related information v Routines for transactional memory
88 XL Fortran: Optimization and Programming Guide TM_SHORT_TRANSACTION_MODE
The TM_SHORT_TRANSACTION_MODE environment variable indicates whether to use the short running speculation mode in the hardware.
NO TM_SHORT_TRANSACTION_MODE = YES
The default value is NO. The value is not case sensitive.
TM_SHORT_TRANSACTION_MODE = YES enables the short running speculation mode. It is recommended if the transactional atomic region is short.
TM_SHORT_TRANSACTION_MODE = NO enables the long running speculation mode. It is recommended if the transactional atomic region is large and many reuses are among the references inside the transaction.
Optimizing your SMP code Most IBM processors are capable of shared-memory parallel processing. Compile with -qsmp to generate the threaded code needed to exploit this capability. The option implies a -O2 optimization level. The default behavior for the option without suboptions is to do automatic parallelization with optimization.
The most commonly used -qsmp suboptions are summarized in the following table. Commonly used -qsmp suboptions Suboption Behavior auto Instructs the compiler to automatically generate parallel code where possible without user assistance. This option also recognizes all the SMP directives. omp Enforces compliance with the OpenMP API for specifying explicit parallelism. opt Instructs the compiler to optimize as well as parallelize. The optimization is equivalent to -O2 –qhot in the absence of other optimization options. The default setting of -qsmp is -qsmp=auto:noomp:opt. suboptions Other values for the suboption provide control over thread scheduling, nested parallelism, locking, and so on.
Developing and running SMP applications v By default, the parallelization performed is both user-directed and automatic. Use -qsmp=omp:noauto if you are compiling an OpenMP program and do not want automatic parallelization. v Before using -qsmp with automatic parallelization, test your programs using optimization and -qhot in a single-threaded manner. v Always use the reentrant compiler invocations (the _r command invocations, like bgxlf_r) when using -qsmp. v By default, the runtime uses all available processors. Do not set the XLSMPOPTS=PARTHDS or OMP_NUM_THREADS variables unless you want to use fewer than the number of available processors. You might want to set the number of executing threads to a small number or to 1 to ease debugging. v If you are using a dedicated machine or node, consider one of the following environment variable settings:
Chapter 7. Parallel programming with XL Fortran 89 – Set BG_SMP_FAST_WAKEUP to YES. – Set OMP_WAIT_POLICY to ACTIVE . – Set SPINS and YIELDS (suboptions of XLSMPOPTS)to0. These settings prevent the operating system from intervening in the scheduling of threads across synchronization boundaries, such as barriers. v When debugging an SMP program, try using -qsmp=noopt (without -O)to make the debugging information produced by the compiler more precise. You can also use the SNAPSHOT directive to create additional program points for storage visibility by flushing registers to memory.
Parallelization directives You can use parallelization directives to exert control over parallelization. This section describes all parallelization directives supported by XL Fortran on Blue Gene/Q platforms. An introduction to parallelization directives These directives allow you to exert control over parallelization. For example, the PARALLEL DO directive specifies that the loop immediately following the directive should be run in parallel. All parallelization directives are comment form directives.
For more information on rules and syntax for comment form directives, see Comment and noncomment form directives in the XL Fortran Language Reference.
XL Fortran supports a number of SMP directives, divided as follows. To ensure the greatest portability of code, OpenMP directives are recommended where possible. Use the OpenMP trigger_constant, $OMP for OpenMP directives, but do not use this trigger_constant with any other directive. OpenMP directives must not appear in PURE and ELEMENTAL procedures. Parallel region construct Parallel constructs form the foundation of OpenMP based parallel execution in XL Fortran. The PARALLEL/END PARALLEL directive pair forms a basic parallel construct. Each time an executing thread enters a parallel region, it creates a team of threads and becomes master of that team. This allows parallel execution to take place within that construct by the threads in that team. The following directives are necessary for a parallel region:
PARALLEL END PARALLEL
Nesting OpenMP, transactional memory, and thread-level speculative execution
This section describes how you can mix parallel regions. The following types of parallel regions can be used without restrictions in the same program if they are not nested: v OpenMP v Transactional memory (TM) v Thread-level speculative execution (SE)
They can also be nested but with some restrictions. The following table describes the behavior of different nesting scenarios.
90 XL Fortran: Optimization and Programming Guide Table 17. Nesting rules for OpenMP, TM, and SE Scenario Description Runtime action BEGIN SE An SE region is The SE nesting is flattened. The inner nested SE BEGIN SE nested inside an region is run speculatively by one thread as part of SE region. the parallel outer SE region. The clauses specified on END SE the inner SE region are still effective. END SE BEGIN SE A TM region is The TM region is run speculatively in SE mode as BEGIN TM nested inside an part of the outer SE region. SE region. END TM END SE BEGIN SE An OpenMP An OpenMP region running in parallel inside the BEGIN OpenMP region is nested speculative SE region causes the SE region to be inside an SE stopped. The stopped SE region is rolled back and run END OpenMP region. nonspeculatively. The inner OpenMP region is run END SE nonspeculatively by multiple threads. BEGIN TM An SE region is The inner SE region is run speculatively in TM mode BEGIN SE nested inside a by one thread as part of the outer TM region. TM region. END SE END TM BEGIN TM A TM region is The TM nesting is flattened. The inner nested TM BEGIN TM nested inside a regions are run speculatively as part of the outer TM TM region. region. END TM END TM BEGIN TM An OpenMP An OpenMP region running in parallel inside the BEGIN OpenMP region is nested speculative TM region causes the TM region to be inside a TM stopped. The stopped transaction is then rolled back END OpenMP region. and run nonspeculatively. The inner OpenMP region END TM is run nonspeculatively by multiple threads. BEGIN OpenMP An SE region is The SE region is run by one thread in nonspeculative BEGIN SE nested inside an mode. The outer OpenMP region is run in parallel. OpenMP region. END SE END OpenMP BEGIN OpenMP A TM region is The outer OpenMP region is run in parallel and the BEGIN TM nested inside an TM region is run speculatively. OpenMP region. END TM END OpenMP BEGIN TM A program The first region is run in TM mode. The second region END TM contains separate is run in SE mode by one thread. TM and SE BEGIN SE regions. If the reset_speculation_mode routine is called after END SE the first TM region, the second SE region is run in parallel speculatively. BEGIN SE A program The first region is run in SE mode. The second region END SE contains separate is run in TM irrevocable mode. SE and TM BEGIN TM regions. If the reset_speculation_mode routine is called after END TM the first SE region, the second TM region is run in parallel speculatively.
Chapter 7. Parallel programming with XL Fortran 91 Work-sharing constructs Work-sharing constructs divide the execution of code enclosed by the construct between threads in a team. For work-sharing to take place, the construct must be enclosed within the dynamic extent of a parallel region. For further information on work-sharing constructs, see the following directives:
DO END DO SECTIONS END SECTIONS WORKSHARE END WORKSHARE SINGLE END SINGLE
Combined parallel work-sharing constructs A combined parallel work-sharing construct allows you to specify a parallel region that already contains a single work-sharing construct. These combined constructs are semantically identical to specifying a parallel construct enclosing a single work-sharing construct. For more information on implementing combined constructs, see the following directives:
PARALLEL DO END PARALLEL DO PARALLEL SECTIONS END PARALLEL SECTIONS PARALLEL WORKSHARE END PARALLEL WORKSHARE
Synchronization constructs The following directives allow you to synchronize the execution of a parallel region by multiple threads in a team:
ATOMIC BARRIER CRITICAL END CRITICAL FLUSH ORDERED END ORDERED TASKWAIT
Other OpenMP directives The following OpenMP directives provide additional SMP functionality:
MASTER END MASTER TASK END TASK THREADPRIVATE
Non-OpenMP SMP directives The following directives provide additional SMP functionality:
DO SERIAL SCHEDULE THREADLOCAL
92 XL Fortran: Optimization and Programming Guide Deprecated directive The SMP directive listed in the following table has been deprecated and might be removed in a future release. Use the corresponding OpenMP directive or clause to obtain the same behavior. Table 18. Deprecated SMP directive SMP directive name OpenMP directive/clause name SCHEDULE SCHEDULE
The following example shows how to replace the deprecated SMP SCHEDULE directive with the OpenMP SCHEDULE clause.
The original code that uses the SMP SCHEDULE directive is as follows: program loop integer, parameter :: N=500 integer :: i !SMP$ SCHEDULE(STATIC) real :: arr(N)
!SMP$ parallel do do i=1, N arr(i) = real(i-1) enddo end program
To obtain the same behavior, you can use the OpenMP SCHEDULE clause, as shown below: program loop integer, parameter :: N=500 integer :: i real :: arr(N)
!$OMP parallel do schedule(static) do i=1, N arr(i) = real(i-1) enddo end program Detailed descriptions of parallelization directives See the alphabetical list of all parallelization directives supported by XL Fortran.
For information on directive clauses, see “Directive clauses” on page 146. ATOMIC Purpose
You can use the ATOMIC directive to access a specific memory location safely within a parallel region. When you use the ATOMIC directive, you ensure that only one thread is accessing the memory location at a time, avoiding errors that might occur from simultaneous reads or writes to the same memory location.
Atomic operations are useful when creating multi-threaded or concurrent algorithms and data structures. Using atomic constructs, you can write more efficient concurrent algorithms with fewer locks.
An atomic construct supports the following kinds of atomic access: v Atomic update
Chapter 7. Parallel programming with XL Fortran 93 Updates the value of a variable atomically. Allows only one thread to write to a shared variable at a time, avoiding errors from simultaneous writes to the same variable. v Atomic read Reads the value of a variable atomically. The value of a shared variable can be read safely, avoiding the danger of reading an intermediate value of the variable when it is accessed simultaneously by a concurrent thread. v Atomic write Writes the value of a variable atomically. The value of a shared variable can be written exclusively to avoid errors from simultaneous writes. v Atomic capture Updates the value of a variable while capturing the original or final value of the variable atomically.
The ATOMIC directive takes effect only if you specify the -qsmp compiler option.
Syntax
Atomic update
UPDATE ATOMIC
atomic_update_statement
END ATOMIC
Atomic read
ATOMIC READ
atomic_capture_statement
END ATOMIC
94 XL Fortran: Optimization and Programming Guide Atomic write
ATOMIC WRITE
atomic_write_statement
END ATOMIC
Atomic capture
ATOMIC CAPTURE
atomic_update_statement
atomic_capture_statement
END ATOMIC
Or
ATOMIC CAPTURE
atomic_capture_statement
atomic_update_statement
END ATOMIC
where atomic_update_statement is one of the following statements: update_variable = update_variable operator expression update_variable = expression operator update_variable update_variable = intrinsic(update_variable, expression_list) update_variable = intrinsic(expression_list, update_variable) atomic_write_statement is: update_variable = expression atomic_capture_statement is: capture_variable = update_variable
Chapter 7. Parallel programming with XL Fortran 95 where: update_variable, capture_variable are both nonpointer, nonallocatable scalar variables of intrinsic type. intrinsic is one of MAX, MIN, IAND, IOR or IEOR. operator is one of +, –, *, /, .AND., .OR., .EQV., .NEQV. or .XOR.. expression is a scalar expression that does not reference update_variable. expression_list is a comma-separated, non-empty list of scalar expressions that do not reference update_variable.
Note: If the intrinsic is IAND, IOR,orIEOR, expression_list can only contain one expression.
Rules
An ATOMIC directive without a clause is equivalent to atomic update, and applies only to the statement that immediately follows it.
All accesses to a certain storage location throughout a concurrent program must be atomic. A non-atomic access to a memory location might break the expected atomic behavior of all atomic accesses to that storage location.
The expression in an atomic statement is not evaluated atomically. You must ensure that no race conditions exist in the calculation.
Within the entire program, if you use the ATOMIC directive to make references to the storage location of an update_variable, all these references must have the same type and type parameters.
capture_variable, expression, and expression_list must not access the same storage location as update_variable.
For atomic capture access, the operation of writing the captured value to the storage location represented by capture_variable is not atomic.
The function intrinsic, the operator operator, and the assignment must be the intrinsic function, operator and assignment, and not a redefined intrinsic function, defined operator or defined assignment.
Examples
Example 1: In this example, multiple threads are updating a counter. ATOMIC is used to ensure that no updates are lost. PROGRAM P R = 0.0 !$OMP PARALLEL DO SHARED(R) DOI=1,10 !$OMP ATOMIC R=R+1.0 END DO PRINT *,R END PROGRAM P
96 XL Fortran: Optimization and Programming Guide Expected output: 10.0
Example 2: In this example, an ATOMIC directive is required, because it is uncertain which element of array Y is updated in each iteration. PROGRAM P INTEGER, DIMENSION(10) :: Y, INDEX INTEGER B
Y=5 READ(*,*) INDEX, B !$OMP PARALLEL DO SHARED(Y) DOI=1,10 !$OMP ATOMIC Y(INDEX(I)) = MIN(Y(INDEX(I)),B) END DO PRINT *, Y END PROGRAM P
Input data: 101088664422 4
Expected output: 5454545454
Example 3: This example demonstrates the usage of atomic capture. FUNCTION fnc(upper) RESULT(ret) INTEGER, INTENT(IN) :: upper INTEGER :: ret INTEGER, SAVE :: iter = 0
!$OMP ATOMIC CAPTURE iter = iter + 1 ret = iter !$OMP END ATOMIC
IF (ret .GT. upper) THEN ret=-1 ENDIF END FUNCTION fnc Related reference: “CRITICAL / END CRITICAL” on page 99 “PARALLEL / END PARALLEL” on page 111
See -qsmp in the Compiler Reference BARRIER Purpose
The BARRIER directive enables you to synchronize all threads in a team. When a thread encounters a BARRIER directive, it will wait until all other threads in the team reach the same point.
Type
The BARRIER directive only takes effect if you specify the -qsmp compiler option.
Chapter 7. Parallel programming with XL Fortran 97 Syntax
BARRIER
Rules
A BARRIER region binds to the closest enclosing PARALLEL region.
A BARRIER region must not be closely nested inside a CRITICAL, MASTER, ORDERED, TASK or work-sharing region.
All threads in the team of the binding parallel region must execute the BARRIER region and complete execution of all explicit tasks in the binding parallel region up to this point before any threads in the team proceed beyond the barrier.
All BARRIER regions and work-sharing region must be encountered in the same order by all threads in the team.
Each BARRIER region must be encountered by all threads in a team or by none at all.
In addition to synchronizing the threads in a team, the BARRIER directive implies the FLUSH directive without the variable_name_list.
Examples
An example of the BARRIER construct binding to the PARALLEL construct. Note: To calculate C, we need to ensure that A and B have been completely assigned to, so threads need to wait. SUBROUTINE SUB1 INTEGER A(1000), B(1000), C(1000) !$OMP PARALLEL !$OMP DO DOI=1,1000 A(I) = SIN(I*2.5) END DO !$OMP END DO NOWAIT !$OMP DO DOJ=1,10000 B(J)=X+COS(J*5.5) END DO !$OMP END DO NOWAIT ... !$OMP BARRIER C=A+B !$OMP END PARALLEL END
98 XL Fortran: Optimization and Programming Guide Related reference: “FLUSH” on page 105 “PARALLEL / END PARALLEL” on page 111
See -qsmp in the Compiler Reference CRITICAL / END CRITICAL Purpose
The CRITICAL construct allows you to define independent blocks of code that are to be run by at most one thread at a time. It includes a CRITICAL directive that is followed by a block of code and ends with an END CRITICAL directive.
Type
The CRITICAL and END CRITICAL directives only take effect if you specify the -qsmp compiler option.
Syntax
CRITICAL ( lock_name )
block
END CRITICAL ( lock_name )
lock_name provides a way of distinguishing different CRITICAL constructs of code. block represents the block of code to be executed by at most one thread at a time.
Rules
The optional lock_name is a name with global scope. You must not use the lock_name to identify any other global entity in the same executable program.
If you specify the lock_name on the CRITICAL directive, you must specify the same lock_name on the corresponding END CRITICAL directive.
If you specify the same lock_name for more than one CRITICAL construct, the compiler will allow only one thread to execute any one of these CRITICAL constructs at any one time. CRITICAL constructs that have different lock_names may be run in parallel.
The same lock protects all CRITICAL constructs that do not have an explicit lock_name. In other words, the compiler will assign the same lock_name, thereby ensuring that only one thread enters any unnamed CRITICAL construct at a time.
Chapter 7. Parallel programming with XL Fortran 99 The lock_name must not share the same name as any local entity of Class 1.
It is illegal to branch into or out of a CRITICAL construct.
The CRITICAL construct may appear anywhere in a program.
Although it is possible to nest a CRITICAL construct within a CRITICAL region, a deadlock situation may result. The -qsmp=rec_locks compiler option can be used to prevent deadlocks. See the XL Fortran Compiler Reference for more information. The OpenMP API does not allow nested CRITICAL regions to have the same name.
CRITICAL and END CRITICAL directives imply the FLUSH directive without the variable_name_list.
Examples
Example 1: This example illustrates the use of a CRITICAL construct to update a shared variable inside a parallel region. The CRITICAL construct restricts only one thread to execute the code at a time. EXPR=0 !$OMP PARALLEL DO PRIVATE (I) DOI=1,100 !$OMP CRITICAL EXPR = EXPR + A(I) * I !$OMP END CRITICAL END DO
Example 2: An example specifying a lock_name on the CRITICAL construct. !$OMP PARALLEL DO PRIVATE(T) DOI=1,100 T = B(I) * B(I-1) !$OMP CRITICAL (LOCK) SUM=SUM+T !$OMP END CRITICAL (LOCK) END DO Related reference: “ATOMIC” on page 93 “FLUSH” on page 105
See Global entity in the Language Reference
See Local entity in the Language Reference “PARALLEL / END PARALLEL” on page 111
See -qsmp in the Compiler Reference DO / END DO Purpose
The DO (work-sharing) construct enables you to divide the execution of the loop among the members of the team that encounter it. The END DO directive enables you to indicate the end of a DO loop that is specified by the DO (work-sharing) directive.
The DO (work-sharing) and END DO directives only take effect when you specify the -qsmp compiler option.
100 XL Fortran: Optimization and Programming Guide Syntax
DO do_clause ,
do_loop
END DO NOWAIT
where do_clause is:
collapse_clause firstprivate_clause lastprivate_clause ordered_clause private_clause reduction_clause schedule_clause
collapse_clause See — “COLLAPSE” on page 148. firstprivate_clause See — “FIRSTPRIVATE” on page 155. lastprivate_clause See — “LASTPRIVATE” on page 157. ordered_clause See — “ORDERED” on page 160 private_clause See — “PRIVATE” on page 161. reduction_clause See — “REDUCTION” on page 164 schedule_clause See — “SCHEDULE” on page 167
Rules
The first noncomment line (not including other directives) that follows the DO (work-sharing) directive must be a DO loop. This line cannot be an infinite DO or DO WHILE loop. The DO (work-sharing) directive applies only to the DO loop that is immediately following the directive, and not to any nested DO loops, unless the COLLAPSE clause is specified.
Chapter 7. Parallel programming with XL Fortran 101 The END DO directive is optional. If you use the END DO directive, it must immediately follow the end of the DO loop.
You may have a DO construct that contains several DO statements. If the DO statements share the same DO termination statement, and an END DO directive follows the construct, you can only specify a work-sharing DO directive for the outermost DO statement of the construct.
If you specify NOWAIT on the END DO directive, a thread that completes its iterations of the loop early will proceed to the instructions following the loop. The thread will not wait for the other threads of the team to complete the DO loop. If you do not specify NOWAIT on the END DO directive, each thread will wait for all other threads within the same team at the end of the DO loop.
If you do not specify the NOWAIT clause, the END DO directive implies the FLUSH directive without the variable_name_list.
All threads in the team must encounter the DO (work-sharing) directive if any thread encounters it. A DO loop must have the same loop boundary and step value for each thread in the team. All work-sharing constructs and BARRIER directives that are encountered must be encountered in the same order by all threads in the team.
A DO (work-sharing) directive must not appear within a CRITICAL, MASTER,or ORDERED region. In addition, it must not appear within a work-sharing region or a TASK region unless it is bound to another parallel region.
You cannot follow a DO (work-sharing) directive by another DO (work-sharing) directive. You can only specify one DO (work-sharing) directive for a given DO loop.
The DO (work-sharing) directive cannot appear with either an INDEPENDENT or DO SERIAL directive for a given DO loop.
To ensure that the same assignment of logical iteration numbers to threads is used in two work-sharing loop regions, you can use the STATIC schedule of the SCHEDULE clause. For details, see “SCHEDULE” on page 167.
Examples
Example 1: An example of several independent DO loops within a PARALLEL construct. No synchronization is performed after the first work-sharing DO loop, because NOWAIT is specified on the END DO directive. !$OMP PARALLEL !$OMP DO DOI=2,N B(I)= (A(I) + A(I-1)) / 2.0 END DO !$OMP END DO NOWAIT !$OMP DO DOJ=2,N C(J) = SQRT(REAL(J*J)) END DO !$OMP END DO C(5) = C(5) + 10 !$OMP END PARALLEL END
102 XL Fortran: Optimization and Programming Guide Example 2: An example of SHARED, and SCHEDULE clauses. !$OMP PARALLEL SHARED(A) !$OMP DO SCHEDULE(STATIC,10) DOI=1,1000 A(I)=I*4 END DO !$OMP END DO !$OMP END PARALLEL
Example 3: An example of both a MASTER and a DO (work-sharing) directive that bind to the closest enclosing PARALLEL directive. !$OMP PARALLEL DEFAULT(PRIVATE), SHARED(X) Y = 100 !$OMP MASTER PRINT *, Y !$OMP END MASTER !$OMP DO DOI=1,10 X(I) = I X(I) = X(I) + Y END DO !$OMP END PARALLEL END
Example 4: An example of both the FIRSTPRIVATE and the LASTPRIVATE clauses on DO (work-sharing) directives. X = 100
!$OMP PARALLEL PRIVATE(I), SHARED(X,Y) !$OMP DO FIRSTPRIVATE(X), LASTPRIVATE(X) DOI=1,80 Y(I)=X+I X=I END DO !$OMP END PARALLEL END Related reference: “COLLAPSE” on page 148
See DO in the Language Reference “DO SERIAL” “FLUSH” on page 105
See INDEPENDENT in the Language Reference “ORDERED / END ORDERED” on page 108 “PARALLEL / END PARALLEL” on page 111 “PARALLEL DO / END PARALLEL DO” on page 113 DO SERIAL Purpose
The DO SERIAL directive indicates to the compiler that the DO loop that is immediately following the directive must not be parallelized. This directive is useful in blocking automatic parallelization for a particular DO loop. The DO SERIAL directive only takes effect if you specify the -qsmp compiler option.
Chapter 7. Parallel programming with XL Fortran 103 Syntax
DO SERIAL
Rules
The first noncomment line (not including other directives) that is following the DO SERIAL directive must be a DO loop. The DO SERIAL directive applies only to the DO loop that immediately follows the directive and not to any loops that are nested within that loop.
You can only specify one DO SERIAL directive for a given DO loop. The DO SERIAL directive must not appear with the DO (work-sharing), or PARALLEL DO directive on the same DO loop.
White space is optional between DO and SERIAL.
You should not use the OpenMP trigger constant with this directive.
Examples
Example 1: An example with nested DO loops where the inner loop (the J loop) is not parallelized. !$OMP PARALLEL DO PRIVATE(S,I), SHARED(A) DO I=1, 500 S=0 !SMP$ DOSERIAL DO J=1, 500 S=S+1 ENDDO A(I)=S+I ENDDO
Example 2: An example with the DOSERIAL directive applied in nested loops. In this case, if automatic parallelization is enabled the I or K loop may be parallelized. DO I=1, 100 !SMP$ DOSERIAL DO J=1, 100 DO K=1, 100 ARR(I,J,K)=I+J+K ENDDO ENDDO ENDDO
104 XL Fortran: Optimization and Programming Guide Related reference: “DO / END DO” on page 100
See DO in the Language Reference “PARALLEL DO / END PARALLEL DO” on page 113
See -qdirective in the Compiler Reference
See -qsmp in the Compiler Reference FLUSH Purpose
The FLUSH directive ensures that each thread has access to data generated by other threads. This directive is required because the compiler may keep values in processor registers if a program is optimized. The FLUSH directive ensures that the memory images that each thread views are consistent.
The FLUSH directive only takes effect if you specify the -qsmp compiler option.
You might be able to improve the performance of your program by using the FLUSH directive instead of the VOLATILE attribute. The VOLATILE attribute causes variables to be flushed after every update and before every use, while FLUSH causes variables to be written to or read from memory only when specified.
Syntax
FLUSH ( variable_name_list )
Rules
You can specify this directive anywhere in your code; however, if you specify it outside a parallel region, it is ignored.
If you specify a variable_name_list, only the variables in that list are written to or read from memory (assuming that they have not been written or read already). All variables in the variable_name_list must be at the current scope and must be thread visible. Thread visible variables can be any of the following: v Globally visible variables (common blocks and module data) v Local and host-associated variables with the SAVE attribute v Local variables without the SAVE attribute that are specified in a SHARED clause in a parallel region within the subprogram v Local variables without the SAVE attribute that have had their addresses taken v All pointer dereferences v Dummy arguments
If an item or a subobject of an item in the variable_name_list has the POINTER attribute, the allocation and association status of the POINTER item is flushed, but the pointer target is not. If an item in the variable_name_list is an integer pointer,
Chapter 7. Parallel programming with XL Fortran 105 the pointer is flushed, but the object to which it points is not. If an item in the variable_name_list has the ALLOCATABLE attribute and the item is allocated, the allocated object is flushed. Otherwise, the allocation status is flushed
If you do not specify a variable_name_list, all thread visible variables are written to or read from memory.
When a thread encounters the FLUSH directive, it writes into memory the modifications to the affected variables. The thread also reads the latest copies of the variables from memory if it has local copies of those variables: for example, if it has copies of the variables in registers.
It is not mandatory for all threads in a team to use the FLUSH directive. However, to guarantee that all thread visible variables are current, any thread that modifies a thread visible variable should use the FLUSH directive to update the value of that variable in memory. If you do not use FLUSH or one of the directives that implies FLUSH (see below), the value of the variable might not be the most recent one.
The FLUSH directive does not imply any ordering between the directive and operations on variables not in the variable_name_list. The FLUSH directive does not imply any ordering between two or more FLUSH constructs if the constructs do not have any variables in common in the variable_name_list.
Note that FLUSH is not atomic. You must FLUSH shared variables that are controlled by a shared lock variable with one directive and then FLUSH the lock variable with another. This guarantees that the shared variables are written before the lock variable.
The following directives imply a FLUSH directive without the variable_name_list unless you specify a NOWAIT clause for those directives to which it applies: v BARRIER v CRITICAL/END CRITICAL v END DO v END SECTIONS v END SINGLE v END WORKSHARE v PARALLEL/END PARALLEL v PARALLEL DO/END PARALLEL DO v PARALLEL SECTIONS/END PARALLEL SECTIONS v PARALLEL WORKSHARE/END PARALLEL WORKSHARE v ORDERED/END ORDERED
The ATOMIC directive implies a FLUSH directive with the variable_name_list. The variable_name_list contains only the object updated in the ATOMIC construct
The following routines imply a FLUSH directive without the variable_name_list: v During OMP_SET_LOCK, and OMP_UNSET_LOCK regions. v During OMP_TEST_LOCK, OMP_SET_NEST_LOCK, OMP_UNSET_NEST_LOCK and OMP_TEST_NEST_LOCK regions, if the region causes the lock to be set or unset.
106 XL Fortran: Optimization and Programming Guide Examples
In the following example, two threads perform calculations in parallel and are synchronized when the calculations are complete: PROGRAM P USE OMP_LIB INTEGER INSYNC(0:1), IAM
!$OMP PARALLEL DEFAULT(PRIVATE) SHARED(INSYNC) NUM_THREADS(2)" IAM = OMP_GET_THREAD_NUM() INSYNC(IAM) = 0 !$OMP BARRIER CALL WORK !$OMP FLUSH(INSYNC) INSYNC(IAM) = 1 ! Each thread sets a flag ! once it has !$OMP FLUSH(INSYNC) ! completed its work. DO WHILE (INSYNC(1-IAM) .eq. 0) ! One thread waits for ! another to complete !$OMP FLUSH(INSYNC) ! its work. END DO
!$OMP END PARALLEL
END PROGRAM P
SUBROUTINE WORK ! Each thread does indep- ! endent calculations. ! ... !$OMP FLUSH ! flush work variables ! before INSYNC ! is flushed. END SUBROUTINE WORK MASTER / END MASTER Purpose
The MASTER construct enables you to define a block of code that will be run by only the master thread of the team. It includes a MASTER directive that precedes a block of code and ends with an END MASTER directive.
The MASTER and END MASTER directives only take effect if you specify the -qsmp compiler option.
Syntax
MASTER
block
END MASTER
block represents the block of code that will be run by the master thread of the team.
Chapter 7. Parallel programming with XL Fortran 107 Rules
It is illegal to branch into or out of a MASTER construct.
A MASTER directive binds to the closest enclosing PARALLEL region, if one exists.
A MASTER directive cannot appear within a work-sharing region or a TASK region.
No implied barrier exists on entry to, or exit from, the MASTER construct.
Examples
Example 1: An example of the MASTER directive binding to the PARALLEL directive. !$OMP PARALLEL DEFAULT(SHARED) !$OMP MASTER Y = 10.0 X = 0.0 DOI=1,4 X=X+COS(Y) + I END DO !$OMP END MASTER !$OMP BARRIER !$OMP DO PRIVATE(J) DOJ=1,10000 A(J)=X+SIN(J*2.5) END DO !$OMP END DO !$OMP END PARALLEL END Related reference:
See -qsmp in the Compiler Reference “PARALLEL / END PARALLEL” on page 111 “DO / END DO” on page 100 ORDERED / END ORDERED Purpose
The ORDERED / END ORDERED directives cause the iterations of a block of code within a parallel loop to be executed in the order that the loop would execute in if it was run sequentially. You can force the code inside the ORDERED construct to run in a predictable order while code outside of the construct runs in parallel.
The ORDERED and END ORDERED directives only take effect if you specify the -qsmp compiler option.
108 XL Fortran: Optimization and Programming Guide Syntax
ORDERED
block
END ORDERED
block represents the block of code that will be executed in sequence.
Rules
The ORDERED directive can only appear in the dynamic extent of a DO or PARALLEL DO directive. It is illegal to branch into or out of an ORDERED construct.
The ORDERED directive binds to the nearest dynamically enclosing DO or PARALLEL DO directive. You must specify the ORDERED clause on the DO or PARALLEL DO directive to which the ORDERED construct binds.
ORDERED constructs that bind to different DO directives are independent of each other.
Only one thread can execute an ORDERED construct at a time. Threads enter the ORDERED construct in the order of the loop iterations. A thread will enter the ORDERED construct if all of the previous iterations have either executed the construct or will never execute the construct.
Each iteration of a parallel loop with an ORDERED construct can only execute that ORDERED construct once. Each iteration of a parallel loop can execute at most one ORDERED directive. An ORDERED construct cannot appear within the dynamic extent of a CRITICAL construct.
The END ORDERED directive implies the FLUSH directive without the variable_name_list
Examples
Example 1: In this example, an ORDERED parallel loop counts down. PROGRAM P !$OMP PARALLEL DO ORDERED DOI=3,1,-1 !$OMP ORDERED CALL C_PRINT(I) ! print I using routine written in C !$OMP END ORDERED END DO END PROGRAM P
The expected output of this program is: 3 2 1
Chapter 7. Parallel programming with XL Fortran 109 Example 2: This example shows a program with two ORDERED constructs in a parallel loop. Each iteration can only execute a single section. PROGRAM P !$OMP PARALLEL DO ORDERED DOI=1,3 IF (MOD(I,2) == 0) THEN !$OMP ORDERED CALL C_PRINT(I*10) ! print I*10 using routine written in C !$OMP END ORDERED ELSE !$OMP ORDERED CALL C_PRINT(I) ! print I using routine written in C !$OMP END ORDERED END IF END DO END PROGRAM P
The expected output of this program is: 1 20 3
Example 3: In this example, the program computes the sum of all elements of an array that are greater than a threshold. ORDERED is used to ensure that the results are always reproducible: roundoff will take place in the same order every time the program is executed, so the program will always produce the same results. PROGRAM P REAL :: A(1000) REAL :: THRESHOLD = 999.9 REAL :: SUM = 0.0
!$OMP PARALLEL DO ORDERED DOI=1,1000 IF (A(I) > THRESHOLD) THEN !$OMP ORDERED SUM = SUM + A(I) !$OMP END ORDERED END IF END DO END PROGRAM P
Note: To avoid bottleneck situations when using the ORDERED clause, you can try using DYNAMIC scheduling or STATIC scheduling with a small chunk size. For more information on scheduling parameters, see the “SCHEDULE” on page 167 clause.
110 XL Fortran: Optimization and Programming Guide Related reference:
See -qsmp in the Compiler Reference “PARALLEL DO / END PARALLEL DO” on page 113 “DO / END DO” on page 100 “CRITICAL / END CRITICAL” on page 99 PARALLEL / END PARALLEL Purpose
The PARALLEL construct enables you to define a block of code that can be executed by a team of threads concurrently. The PARALLEL construct includes a PARALLEL directive that is followed by one or more blocks of code, and ends with an END PARALLEL directive.
The PARALLEL and END PARALLEL directives only take effect if you specify the -qsmp compiler option.
Syntax
PARALLEL parallel_clause ,
block
END PARALLEL
where parallel_clause is:
copyin_clause default_clause firstprivate_clause IF ( scalar_logical_expr ) num_threads_clause private_clause reduction_clause shared_clause copyin_clause See — “COPYIN” on page 150 default_clause See — “DEFAULT” on page 152 if_clause See — “IF” on page 156 firstprivate_clause See — “FIRSTPRIVATE” on page 155.
Chapter 7. Parallel programming with XL Fortran 111 num_threads_clause See — “NUM_THREADS” on page 160. private_clause See — “PRIVATE” on page 161. reduction_clause See — “REDUCTION” on page 164 shared_clause See — “SHARED” on page 170
Rules
It is illegal to branch into or out of a PARALLEL construct.
The IF and DEFAULT clauses can appear at most once in a PARALLEL directive.
You should be careful when you perform input/output operations in a parallel region. If multiple threads execute a Fortran I/O statement on the same unit, you should make sure that the threads are synchronized. If you do not, the behavior is undefined. See “Parallel I/O issues” on page 281 for more information. Also note that although in the XL Fortran implementation each thread has exclusive access to the I/O unit, the OpenMP specification does not require exclusive access.
Directives that bind to a parallel region will bind to that parallel region even if it is serialized.
The END PARALLEL directive implies the FLUSH directive without the variable_name_list and a BARRIER directive.
Examples
Example 1: An example of an outer PARALLEL directive with the PRIVATE clause enclosing the PARALLEL construct. Note: The SHARED clause is present on the inner PARALLEL construct. !$OMP PARALLEL PRIVATE(X) !$OMP DO DOI=1,10 X(I) = I !$OMP PARALLEL SHARED (X,Y) !$OMP DO DOK=1,10 Y(K,I)= K * X(I) END DO !$OMP END DO !$OMP END PARALLEL END DO !$OMP END DO !$OMP END PARALLEL
Example 2: This example demonstrates the use of the COPYIN clause. Each thread created by the PARALLEL directive has its own copy of the common block BLOCK. The COPYIN clause causes the initial value of FCTR to be copied into the threads that execute iterations of the DO loop. PROGRAM TT COMMON /BLOCK/ FCTR INTEGER :: I, FCTR !$OMP THREADPRIVATE(/BLOCK/) INTEGER :: A(100)
112 XL Fortran: Optimization and Programming Guide FCTR = -1 A=0
!$OMP PARALLEL COPYIN(FCTR) !$OMP DO DO I=1, 100 FCTR = FCTR + I CALL SUB(A(I), I) ENDDO !$OMP END PARALLEL
PRINT *, A END PROGRAM
SUBROUTINE SUB(AA, J) INTEGER :: FCTR, AA, J COMMON /BLOCK/ FCTR !$OMP THREADPRIVATE(/BLOCK/) ! EACH THREAD GETS ITS OWN COPY ! OF BLOCK. AA = FCTR FCTR = FCTR - J END SUBROUTINE SUB
The expected output is: 0123...96979899 Related reference: “FLUSH” on page 105 “PARALLEL DO / END PARALLEL DO”
See INDEPENDENT in the Language Reference “THREADPRIVATE” on page 138 “DO / END DO” on page 100
See -qdirective in the Compiler Reference
See -qsmp in the Compiler Reference PARALLEL DO / END PARALLEL DO Purpose
The PARALLEL DO directive enables you to specify which loops the compiler should parallelize. This is semantically equivalent to: !$OMP PARALLEL !$OMP DO ... !$OMP ENDDO !$OMP END PARALLEL and is a convenient way of parallelizing loops. The END PARALLEL DO directive allows you to indicate the end of a DO loop that is specified by the PARALLEL DO directive.
The PARALLEL DO and END PARALLEL DO directives only take effect if you specify the -qsmp compiler option.
Chapter 7. Parallel programming with XL Fortran 113 Syntax
PARALLEL DO parallel_do_clause ,
parallel_do_loop
END PARALLEL DO
where parallel_do_clause is:
collapse_clause copyin_clause default_clause firstprivate_clause IF ( scalar_logical_expr ) lastprivate_clause num_threads_clause ordered_clause private_clause reduction_clause SCHEDULE ( sched_type ) ,n shared_clause
collapse_clause See — “COLLAPSE” on page 148 copyin_clause See — “COPYIN” on page 150 default_clause See — “DEFAULT” on page 152 if_clause See — “IF” on page 156. firstprivate_clause See — “FIRSTPRIVATE” on page 155. lastprivate_clause See — “LASTPRIVATE” on page 157. num_threads_clause See — “NUM_THREADS” on page 160. ordered_clause See — “ORDERED” on page 160 private_clause See — “PRIVATE” on page 161
114 XL Fortran: Optimization and Programming Guide reduction_clause See — “REDUCTION” on page 164 schedule_clause See — “SCHEDULE” on page 167 shared_clause See — “SHARED” on page 170
Rules
The first noncomment line (not including other directives) that is following the PARALLEL DO directive must be a DO loop. This line cannot be an infinite DO or DO WHILE loop. The PARALLEL DO directive applies only to the DO loop that is immediately following the directive, and not to any nested DO loops, unless the COLLAPSE clause is specified.
If you specify a DO loop by a PARALLEL DO directive, the END PARALLEL DO directive is optional. If you use the END PARALLEL DO directive, it must immediately follow the end of the DO loop.
You may have a DO construct that contains several DO statements. If the DO statements share the same DO termination statement, and an END PARALLEL DO directive follows the construct, you can only specify a PARALLEL DO directive for the outermost DO statement of the construct.
You must not follow the PARALLEL DO directive by a DO (work-sharing) or DO SERIAL directive. You can specify only one PARALLEL DO directive for a given DO loop.
All work-sharing constructs and BARRIER directives that are encountered must be encountered in the same order by all threads in the team.
The PARALLEL DO directive must not appear with the INDEPENDENT directive for a given DO loop.
Note: You should use the PARALLEL DO directive for maximum portability across multiple vendors. The PARALLEL DO directive is a prescriptive directive, while the INDEPENDENT directive is an assertion about the characteristics of the loop. (See the INDEPENDENT directive in the XL Fortran Language Reference for more information.)
The IF clause may appear at most once in a PARALLEL DO directive.
An IF expression is evaluated outside of the context of the parallel construct. Any function reference in the IF expression must not have side effects.
By default, a nested parallel loop is serialized, regardless of the setting of the IF clause. You can change this default by using the -qsmp=nested_par compiler option.
If the REDUCTION variable of an inner DO loop appears in the PRIVATE or LASTPRIVATE clause of an enclosing DO loop or PARALLEL SECTIONS construct, the variable must be initialized before the inner DO loop.
Chapter 7. Parallel programming with XL Fortran 115 A variable that appears in the REDUCTION clause of an INDEPENDENT directive of an enclosing DO loop must not also appear in the data_scope_entity_list of the PRIVATE or LASTPRIVATE clause.
Within a PARALLEL DO construct, variables that do not appear in the PRIVATE clause are assumed to be shared by default.
You should be careful when you perform input/output operations in a parallel region. If multiple threads execute a Fortran I/O statement on the same unit, you should make sure that the threads are synchronized. If you do not, the behavior is undefined. Also note that although in the XL Fortran implementation each thread has exclusive access to the I/O unit, the OpenMP specification does not require exclusive access.
Directives that bind to a parallel region will bind to that parallel region even if it is serialized.
The END PARALLEL DO directive implies the FLUSH directive without the variable_name_list and a BARRIER directive.
Examples
Example 1: A valid example with the LASTPRIVATE clause. !$OMP PARALLEL DO PRIVATE(I), LASTPRIVATE (X) DO I = 1,10 X=I*I A(I)=X*B(I) END DO PRINT *, X ! X has the value 100
Example 2: A valid example with the REDUCTION clause. !$OMP PARALLEL DO PRIVATE(I), REDUCTION(+:MYSUM) DOI=1,10 MYSUM = MYSUM + IARR(I) END DO
Example 3: A valid example where more than one thread accesses a variable that is marked as SHARED, but the variable is used only in a CRITICAL construct. !$OMP PARALLEL DO SHARED (X) DOI=1,10 A(I) = A(I) * I !$OMP CRITICAL X=X+A(I) !$OMP END CRITICAL END DO
Example 4: A valid example of the END PARALLEL DO directive. REAL A(100), B(2:100), C(100) !$OMP PARALLEL DO DOI=2,100 B(I) = (A(I) + A(I-1))/2.0 END DO !$OMP END PARALLEL DO !$OMP PARALLEL DO DO J = 1, 100 C(J)=X+COS(J*5.5) END DO !$OMP END PARALLEL DO END
116 XL Fortran: Optimization and Programming Guide Related reference: “COLLAPSE” on page 148
See -qdirective in the Compiler Reference
See -qsmp in the Compiler Reference
See DO in the Language Reference “DO / END DO” on page 100
See INDEPENDENT in the Language Reference “ORDERED / END ORDERED” on page 108 “PARALLEL / END PARALLEL” on page 111 “PARALLEL SECTIONS / END PARALLEL SECTIONS” “SCHEDULE” on page 120 “THREADPRIVATE” on page 138 PARALLEL SECTIONS / END PARALLEL SECTIONS Purpose
The PARALLEL SECTIONS construct provides a short form method for including SECTIONS directive inside a PARALLEL construct.
The PARALLEL SECTIONS, SECTION and END PARALLEL SECTIONS directives only take effect if you specify the -qsmp compiler option.
Syntax
PARALLEL SECTIONS parallel_sections_clause ,
block SECTION SECTION block
END PARALLEL SECTIONS
where parallel_sections_clause is:
Chapter 7. Parallel programming with XL Fortran 117 copyin_clause default_clause firstprivate_clause IF ( scalar_logical_expr ) lastprivate_clause num_threads_clause private_clause reduction_clause shared_clause
copyin_clause See — “COPYIN” on page 150 default_clause See — “DEFAULT” on page 152 firstprivate_clause See — “FIRSTPRIVATE” on page 155. if_clause See — “IF” on page 156 lastprivate_clause See — “LASTPRIVATE” on page 157. num_threads_clause See — “NUM_THREADS” on page 160. private_clause See — “PRIVATE” on page 161. reduction_clause See — “REDUCTION” on page 164 shared_clause See — “SHARED” on page 170
Rules
See the Rules section in “SECTIONS / END SECTIONS” on page 123.
In a PARALLEL SECTIONS construct, a variable that appears in the REDUCTION clause of an INDEPENDENT directive or the PARALLEL DO directive of an enclosing DO loop must not also appear in the data_scope_entity_list of the PRIVATE clause.
If the REDUCTION variable of the inner PARALLEL SECTIONS construct appears in the PRIVATE clause of an enclosing DO loop or PARALLEL SECTIONS construct, the variable must be initialized before the inner PARALLEL SECTIONS construct.
Examples
Example 1: !$OMP PARALLEL SECTIONS !$OMP SECTION DOI=1,10 C(I) = MAX(A(I),A(I+1)) END DO
118 XL Fortran: Optimization and Programming Guide !$OMP SECTION W=U+V Z=X+Y !$OMP END PARALLEL SECTIONS
Example 2: In this example, the index variable I is declared as PRIVATE. Note also that the first optional SECTION directive has been omitted. !$OMP PARALLEL SECTIONS PRIVATE(I) DOI=1,100 A(I) = A(I) * I END DO !$OMP SECTION CALL NORMALIZE (B) DOI=1,100 B(I) = B(I) + 1.0 END DO !$OMP SECTION DOI=1,100 C(I) = C(I) * C(I) END DO !$OMP END PARALLEL SECTIONS Related reference: “PARALLEL / END PARALLEL” on page 111 “SECTIONS / END SECTIONS” on page 123
See INDEPENDENT in the Language Reference
See -qdirective in the Compiler Reference
See -qsmp in the Compiler Reference PARALLEL WORKSHARE / END PARALLEL WORKSHARE Purpose
The PARALLEL WORKSHARE construct provides a short form method for including a WORKSHARE directive inside a PARALLEL construct.
The PARALLEL WORKSHARE / END PARALLEL WORKSHARE directives only take effect if you specify the -qsmp compiler option
Syntax
PARALLEL WORKSHARE parallel_workshare_clause ,
block
END PARALLEL WORKSHARE
where parallel_workshare_clause is any of the directives accepted by either the PARALLEL or WORKSHARE directives.
Chapter 7. Parallel programming with XL Fortran 119 Related reference: “PARALLEL / END PARALLEL” on page 111 “WORKSHARE / END WORKSHARE” on page 144 SCHEDULE Purpose
Note: The SCHEDULE directive has been deprecated and might be removed in a future release. Use the corresponding OpenMP SCHEDULE clause. For more information about the deprecated SMP directives and deprecation examples, see “Deprecated directive” on page 93.
The SCHEDULE directive allows the user to specify the chunking method for parallelization. Work is assigned to threads in different manners depending on the scheduling type or chunk size used.
The SCHEDULE directive only takes effect if you specify the -qsmp compiler option.
Syntax
SCHEDULE ( sched_type ) , n
nnmust be a positive, specification expression. You must not specify n for the sched_type RUNTIME. sched_type is AFFINITY, DYNAMIC, GUIDED, RUNTIME,orSTATIC
For more information on sched_type parameters, see the SCHEDULE clause. number_of_iterations is the number of iterations in the loop to be parallelized. number_of_threads is the number of threads used by the program.
Rules
The SCHEDULE directive must appear in the specification part of a scoping unit.
Only one SCHEDULE directive may appear in the specification part of a scoping unit.
The SCHEDULE directive applies to the situation when all loops in the scoping unit do not already have explicit scheduling types specified. Individual loops can have scheduling types specified using the SCHEDULE clause of the PARALLEL DO directive.
Any dummy arguments appearing or referenced in the specification expression for the chunk size n must also appear in the SUBROUTINE or FUNCTION statement and in all ENTRY statements appearing in the given subprogram.
If the specified chunk size n is greater than the number of iterations, the loop will not be parallelized and will execute on a single thread.
120 XL Fortran: Optimization and Programming Guide If you specify more than one method of determining the chunking algorithm, the compiler will follow, in order of precedence: 1. SCHEDULE clause to the PARALLEL DO directive. 2. SCHEDULE directive. 3. schedule suboption to the -qsmp compiler option. See the -qsmp option in the XL Fortran Compiler Reference. 4. XLSMPOPTS runtime option. See “XLSMPOPTS” on page 74. 5. runtime default (that is, STATIC).
Examples
Example 1. Given the following information: number of iterations = 1000 number of threads = 4 and using the GUIDED scheduling type, the chunk sizes would be as follows: 25018814110679594533251914118643321111
The iterations would then be divided into the following chunks: chunk 1 = iterations 1 to 250 chunk 2 = iterations 251 to 438 chunk 3 = iterations 439 to 579 chunk 4 = iterations 580 to 685 chunk 5 = iterations 686 to 764 chunk 6 = iterations 765 to 823 chunk 7 = iterations 824 to 868 chunk 8 = iterations 869 to 901 chunk 9 = iterations 902 to 926 chunk 10 = iterations 927 to 945 chunk 11 = iterations 946 to 959 chunk 12 = iterations 960 to 970 chunk 13 = iterations 971 to 978 chunk 14 = iterations 979 to 984 chunk 15 = iterations 985 to 988 chunk 16 = iterations 989 to 991 chunk 17 = iterations 992 to 994 chunk 18 = iterations 995 to 996 chunk 19 = iterations 997 to 997 chunk 20 = iterations 998 to 998 chunk 21 = iterations 999 to 999 chunk 22 = iterations 1000 to 1000
A possible scenario for the division of work could be: thread 1 executes chunks1510131820 thread 2 executes chunks 2 7 9 14 16 22 thread 3 executes chunks36121519 thread 4 executes chunks48111721
Example 2. Given the following information: number of iterations = 100 number of threads = 4 and using the AFFINITY scheduling type, the iterations would be divided into the following partitions: partition 1 = iterations 1 to 25 partition 2 = iterations 26 to 50 partition 3 = iterations 51 to 75 partition 4 = iterations 76 to 100
Chapter 7. Parallel programming with XL Fortran 121 The partitions would be divided into the following chunks: chunk 1a = iterations 1 to 13 chunk 1b = iterations 14 to 19 chunk 1c = iterations 20 to 22 chunk 1d = iterations 23 to 24 chunk 1e = iterations 25 to 25
chunk 2a = iterations 26 to 38 chunk 2b = iterations 39 to 44 chunk 2c = iterations 45 to 47 chunk 2d = iterations 48 to 49 chunk 2e = iterations 50 to 50
chunk 3a = iterations 51 to 63 chunk 3b = iterations 64 to 69 chunk 3c = iterations 70 to 72 chunk 3d = iterations 73 to 74 chunk 3e = iterations 75 to 75
chunk 4a = iterations 76 to 88 chunk 4b = iterations 89 to 94 chunk 4c = iterations 95 to 97 chunk 4d = iterations 98 to 99 chunk 4e = iterations 100 to 100
A possible scenario for the division of work could be: thread 1 executes chunks 1a 1b 1c 1d 1e 4d thread 2 executes chunks 2a 2b 2c 2d thread 3 executes chunks 3a 3b 3c 3d 3e 2e thread 4 executes chunks 4a 4b 4c 4e
In this scenario, thread 1 finished executing all the chunks in its partition and then grabbed an available chunk from the partition of thread 4. Similarly, thread 3 finished executing all the chunks in its partition and then grabbed an available chunk from the partition of thread 2.
Example 3. Given the following information: number of iterations = 1000 number of threads = 4
and using the DYNAMIC scheduling type and chunk size of 100, the chunk sizes would be as follows: 100 100 100 100 100 100 100 100 100 100
The iterations would be divided into the following chunks: chunk 1 = iterations 1 to 100 chunk 2 = iterations 101 to 200 chunk 3 = iterations 201 to 300 chunk 4 = iterations 301 to 400 chunk 5 = iterations 401 to 500 chunk 6 = iterations 501 to 600 chunk 7 = iterations 601 to 700 chunk 8 = iterations 701 to 800 chunk 9 = iterations 801 to 900 chunk 10 = iterations 901 to 1000
A possible scenario for the division of work could be: thread 1 executes chunks 1 5 9 thread 2 executes chunks 2 8 thread 3 executes chunks 3 6 10 thread 4 executes chunks 4 7
122 XL Fortran: Optimization and Programming Guide Example 4. Given the following information: number of iterations = 100 number of threads = 4 and using the STATIC scheduling type, the iterations would be divided into the following chunks: chunk 1 = iterations 1 to 25 chunk 2 = iterations 26 to 50 chunk 3 = iterations 51 to 75 chunk 4 = iterations 76 to 100
A possible scenario for the division of work could be: thread 1 executes chunks 1 thread 2 executes chunks 2 thread 3 executes chunks 3 thread 4 executes chunks 4 Related reference:
See DO in the Language Reference SECTIONS / END SECTIONS Purpose
The SECTIONS construct defines distinct blocks of code to be executed in parallel by threads in the team.
The SECTIONS and END SECTIONS directives only take effect if you specify the -qsmp compiler option.
Syntax
SECTIONS sections_clause ,
block SECTION SECTION block
END SECTIONS NOWAIT
where sections_clause is:
firstprivate_clause lastprivate_clause private_clause reduction_clause
Chapter 7. Parallel programming with XL Fortran 123 firstprivate_clause See — “FIRSTPRIVATE” on page 155. lastprivate_clause See — “LASTPRIVATE” on page 157. private_clause See — “PRIVATE” on page 161. reduction_clause See — “REDUCTION” on page 164
Rules
The SECTIONS construct must be encountered by all threads in a team or by none of the threads in a team. All work-sharing constructs and BARRIER directives that are encountered must be encountered in the same order by all threads in the team.
The SECTIONS construct includes the delimiting directives, and the blocks of code they enclose. At least one block of code must appear in the construct.
You must specify the SECTION directive at the beginning of each block of code except for the first. The end of a block is delimited by either another SECTION directive or by the END SECTIONS directive.
It is illegal to branch into or out of any block of code that is enclosed in the SECTIONS construct. All SECTION directives must appear within the lexical extent of the SECTIONS/END SECTIONS directive pair.
The scheduling of structured blocks among threads in the team is set so that the first thread arriving is the first thread to execute the block. The compiler determines how to divide the work among the threads based on a number of factors, such as the number of threads in the team and the number of sections to be executed in parallel. In a SECTIONS construct, a single thread might execute more than one SECTION. It is also possible that a thread in the team might not execute any SECTION.
In order for the directive to execute in parallel, you must place the SECTIONS/END SECTIONS pair within a parallel region. Otherwise, the blocks will be executed serially.
If you specify NOWAIT on the SECTIONS directive, a thread that completes its sections early will proceed to the instructions following the SECTIONS construct. If you do not specify the NOWAIT clause, each thread will wait for all of the other threads in the same team to reach the END SECTIONS directive. However, there is no implied BARRIER at the start of the SECTIONS construct.
You cannot specify a SECTIONS directive within the dynamic extent of a CRITICAL, MASTER, ORDERED,orTASK directive.
You cannot nest SECTIONS, DO or SINGLE directives that bind to the same PARALLEL directive.
BARRIER and MASTER directives are not permitted in the dynamic extent of a SECTIONS directive.
The END SECTIONS directive implies the FLUSH directive.
124 XL Fortran: Optimization and Programming Guide Examples
Example 1: This example shows a valid use of the SECTIONS construct within a PARALLEL region. INTEGER :: I, B(500), S, SUM ! ... S=0 SUM=0 !$OMP PARALLEL SHARED(SUM), FIRSTPRIVATE(S) !$OMP SECTIONS REDUCTION(+: SUM), LASTPRIVATE(I) !$OMP SECTION S = FCT1(B(1::2)) ! Array B is not altered in FCT1. SUM=SUM+S ! ... !$OMP SECTION S = FCT2(B(2::2)) ! Array B is not altered in FCT2. SUM=SUM+S ! ... !$OMP SECTION DOI=1,500 !Thelocal copy of S is initialized S=S+B(I) ! to zero. END DO SUM=SUM+S ! ... !$OMP END SECTIONS ! ... !$OMP DO REDUCTION(-: SUM) DO J=I-1, 1, -1 ! The loop starts at 500 -- the last ! value from the previous loop. SUM = SUM - B(J) END DO
!$OMP MASTER SUM = SUM - FCT1(B(1::2)) - FCT2(B(2::2)) !$OMP END MASTER !$OMP END PARALLEL ! ... ! Upon termination of the PARALLEL ! region, the value of SUM remains zero.
Example 2: This example shows a valid use of nested SECTIONS. !$OMP PARALLEL !$OMP MASTER CALL RANDOM_NUMBER(CX) CALL RANDOM_NUMBER(CY) CALL RANDOM_NUMBER(CZ) !$OMP END MASTER
!$OMP SECTIONS !$OMP SECTION !$OMP PARALLEL !$OMP SECTIONS PRIVATE(I) !$OMP SECTION DO I=1, 5000 X(I) = X(I) + CX END DO !$OMP SECTION DO I=1, 5000 Y(I) = Y(I) + CY END DO !$OMP END SECTIONS !$OMP END PARALLEL
!$OMP SECTION !$OMP PARALLEL SHARED(CZ,Z)
Chapter 7. Parallel programming with XL Fortran 125 !$OMP DO DO I=1, 5000 Z(I) = Z(I) + CZ END DO !$OMP END DO !$OMP END PARALLEL !$OMP END SECTIONS NOWAIT
! The following computations do not ! depend on the results from the ! previous section.
!$OMP DO DO I=1, 5000 T(I) = T(I) * CT END DO !$OMP END DO !$OMP END PARALLEL Related reference: “PARALLEL / END PARALLEL” on page 111 “BARRIER” on page 97 “PARALLEL DO / END PARALLEL DO” on page 113
See INDEPENDENT in the Language Reference “THREADPRIVATE” on page 138
See -qdirective in the Compiler Reference
See -qsmp in the Compiler Reference SINGLE / END SINGLE Purpose
You can use the SINGLE / END SINGLE directive construct to specify that the enclosed code should only be executed by one thread in the team.
The SINGLE directive only takes effect if you specify the –qsmp compiler option.
Syntax
SINGLE single_clause ,
block
END SINGLE NOWAIT end_single_clause
126 XL Fortran: Optimization and Programming Guide where single_clause is:
private_clause firstprivate_clause private_clause See — “PRIVATE” on page 161. firstprivate_clause See — “FIRSTPRIVATE” on page 155. where end_single_clause is:
copyprivate_clause ,
NOWAIT copyprivate_clause See — “COPYPRIVATE” on page 151.
Rules
It is illegal to branch into or out of a block that is enclosed within the SINGLE construct.
The SINGLE construct must be encountered by all threads in a team or by none of the threads in a team. The first thread to encounter the SINGLE construct will execute it. All work-sharing constructs and BARRIER directives that are encountered must be encountered in the same order by all threads in the team.
If you specify NOWAIT on the END SINGLE directive, the threads that are not executing the SINGLE construct will proceed to the instructions following the SINGLE construct. If you do not specify the NOWAIT clause, each thread will wait at the END SINGLE directive until the thread executing the construct reaches the END SINGLE directive. You may not specify NOWAIT and COPYPRIVATE as part of the same END SINGLE directive.
There is no implied BARRIER at the start of the SINGLE construct. If you do not specify the NOWAIT clause, the BARRIER directive is implied at the END SINGLE directive.
You cannot nest work-sharing constructs inside one another if they bind to the same PARALLEL directive.
SINGLE directives are not permitted within the CRITICAL, MASTER, ORDERED,orTASK regions. BARRIER and MASTER directives are not permitted within the SINGLE regions.
If you have specified a variable as PRIVATE, FIRSTPRIVATE, LASTPRIVATE,or REDUCTION in the PARALLEL construct which encloses your SINGLE construct, you cannot specify the same variable in the PRIVATE or FIRSTPRIVATE clause of the SINGLE construct.
Chapter 7. Parallel programming with XL Fortran 127 The SINGLE directive binds to the closest enclosing PARALLEL region, if one exists.
Examples
Example 1: In this example, the BARRIER directive is used to ensure that all threads finish their work before entering the SINGLE construct. REAL :: X(100), Y(50) ! ... !$OMP PARALLEL DEFAULT(SHARED) CALL WORK(X)
!$OMP BARRIER !$OMP SINGLE CALL OUTPUT(X) CALL INPUT(Y) !$OMP END SINGLE
CALL WORK(Y) !$OMP END PARALLEL
Example 2: In this example, the SINGLE construct ensures that only one thread is executing a block of code. In this case, array B is initialized in the DO (work-sharing) construct. After the initialization, a single thread is employed to perform the summation. INTEGER :: I, J REAL :: B(500,500), SM ! ...
J = ... SM = 0.0 !$OMP PARALLEL !$OMP DO PRIVATE(I) DO I=1, 500 CALL INITARR(B(I,:), I) ! initialize the array B ENDDO !$OMP END DO
!$OMP SINGLE ! employ only one thread DO I=1, 500 SM = SM + SUM(B(J:J+1,I)) ENDDO !$OMP END SINGLE
!$OMP DO PRIVATE(I) DO I=500, 1, -1 CALL INITARR(B(I,:), 501-I) ! re-initialize the array B ENDDO !$OMP END PARALLEL
Example 3: This example shows a valid use of the PRIVATE clause. Array X is PRIVATE to the SINGLE construct. If you were to reference array X immediately following the construct, it would be undefined. REAL :: X(2000), A(1000), B(1000)
!$OMP PARALLEL ! ... !$OMP SINGLE PRIVATE(X) CALL READ_IN_DATA(X) A = X(1::2)
128 XL Fortran: Optimization and Programming Guide B = X(2::2) !$OMP END SINGLE ! ... !$OMP END PARALLEL
Example 4: In this example, the LASTPRIVATE variable I is used in allocating TMP, the PRIVATE variable in the SINGLE construct. SUBROUTINE ADD(A, UPPERBOUND) INTEGER :: A(UPPERBOUND), I, UPPERBOUND INTEGER, ALLOCATABLE :: TMP(:) ! ... !$OMP PARALLEL !$OMP DO LASTPRIVATE(I) DO I=1, UPPERBOUND A(I)=I+1 ENDDO !$OMP END DO
!$OMP SINGLE FIRSTPRIVATE(I), PRIVATE(TMP) ALLOCATE(TMP(0:I-1)) TMP = (/ (A(J),J=I,1,-1) /) ! ... DEALLOCATE(TMP) !$OMP END SINGLE !$OMP END PARALLEL ! ... END SUBROUTINE ADD
Example 5: In this example, a value for the variable I is entered by the user. This value is then copied into the corresponding variable I for all other threads in the team using a COPYPRIVATE clause on an END SINGLE directive. INTEGER I !$OMP PARALLEL PRIVATE (I) ! ... !$OMP SINGLE READ (*, *) I !$OMP END SINGLE COPYPRIVATE (I) ! In all threads in the team, I ! is equal to the value ! ... ! that you entered. !$OMP END PARALLEL
Example 6: In this example, variable J with a POINTER attribute is specified in a COPYPRIVATE clause on an END SINGLE directive. The value of J, not the value of the object that it points to, is copied into the corresponding variable J for all other threads in the team. The object itself is shared among all the threads in the team. INTEGER, POINTER :: J !$OMP PARALLEL PRIVATE (J) ! ... !$OMP SINGLE ALLOCATE (J) READ (*, *) J !$OMP END SINGLE COPYPRIVATE (J) !$OMP ATOMIC J=J+OMP_GET_THREAD_NUM() !$OMP BARRIER !$OMP SINGLE WRITE (*, *) ’J = ’, J ! The result is the sum of all values added to ! J. This result shows that the pointer object ! is shared by all threads in the team. DEALLOCATE (J) !$OMP END SINGLE !$OMP END PARALLEL
Chapter 7. Parallel programming with XL Fortran 129 Related reference: “BARRIER” on page 97 “CRITICAL / END CRITICAL” on page 99 “FLUSH” on page 105 “MASTER / END MASTER” on page 107 “PARALLEL / END PARALLEL” on page 111 SPECULATIVE DO / END SPECULATIVE DO
The SPECULATIVE DO directive instructs the compiler to speculatively parallelize a DO loop.
Syntax
,
!SEP$ SPECULATIVE DO speculative_do_clause
!SEP$ END SPECULATIVE DO
where speculative_do_clause is:
default_clause private_clause firstprivate_clause lastprivate_clause shared_clause num_threads_clause reduction_clause schedule_clause
speculative_do_clause is any of the following clauses: default_clause For details, see “DEFAULT” on page 152. private_clause For details, see “PRIVATE” on page 161. firstprivate_clause For details, see “FIRSTPRIVATE” on page 155. lastprivate_clause For details, see “LASTPRIVATE” on page 157. shared_clause For details, see “SHARED” on page 170. num_threads_clause For details, see “NUM_THREADS” on page 160. reduction_clause For details, see “REDUCTION” on page 164.
130 XL Fortran: Optimization and Programming Guide schedule_clause For details, see “SCHEDULE” on page 167.
Usage
The directives for thread-level speculative execution only take effect if you specify the -qsmp=speculative compiler option.
The SPECULATIVE DO directive precedes a DO loop. The optional END SPECULATIVE DO directive indicates the end of the DO loop.
Example INTEGER ::p=0 INTEGER ::i=1
!SEP$ SPECULATIVE DO FIRSTPRIVATE(p) DOi=1,10 p=p+1 END DO
END
Related information v The "-qsmp" compiler option v Thread-level speculative execution v SPECULATIVE SECTIONS v Routines for thread-level speculative execution v Environment variables for thread-level speculative execution SPECULATIVE SECTIONS / END SPECULATIVE SECTIONS
The SPECULATIVE SECTIONS directive instructs the compiler to speculatively parallelize sections of the code. In code blocks delimited by SPECULATIVE SECTIONS and END SPECULATIVE SECTIONS, you can use the SPECULATIVE SECTION directive to delimit program code segments.
Syntax
,
!SEP$ SPECULATIVE SECTIONS speculative_section_clause
structured_block !SEP$ SPECULATIVE SECTION
!SEP$ SPECULATIVE SECTION structured_block
Chapter 7. Parallel programming with XL Fortran 131 !SEP$ END SPECULATIVE SECTIONS
where speculative_section_clause is:
default_clause private_clause firstprivate_clause lastprivate_clause shared_clause num_threads_clause reduction_clause
speculative_section_clause is any of the following clauses: default_clause For details, see “DEFAULT” on page 152. private_clause For details, see “PRIVATE” on page 161. firstprivate_clause For details, see “FIRSTPRIVATE” on page 155. lastprivate_clause For details, see “LASTPRIVATE” on page 157. shared_clause For details, see “SHARED” on page 170. num_threads_clause For details, see “NUM_THREADS” on page 160. reduction_clause For details, see “REDUCTION” on page 164.
Usage
The directives for thread-level speculative execution only take effect if you specify the "-qsmp=speculative" compiler option.
The SPECULATIVE SECTION directive is optional for the first code segment inside the SPECULATIVE SECTIONS directive. The code segments after the first one must be preceded by a SPECULATIVE SECTION directive. You must use all the SPECULATIVE SECTION directives only in the lexical construct of the code segment that is delimited by the SPECULATIVE SECTIONS and END SPECULATIVE SECTIONS directives.
Example INTEGER ::p=0 INTEGER ::i=1
!SEP$ SPECULATIVE SECTIONS FIRSTPRIVATE(p) p=p+1 !SEP$ SPECULATIVE SECTION
132 XL Fortran: Optimization and Programming Guide i=i-1 !SEP$ END SPECULATIVE SECTIONS
END
Related information v The "-qsmp" compiler option v Thread-level speculative execution v SPECULATIVE DO / END SPECULATIVE DO v Routines for thread-level speculative execution v Environment variables for thread-level speculative execution TASK / END TASK Purpose
The TASK directive instructs the compiler to run a block of code in parallel with the code outside the task region. The TASK directive can be useful for parallelizing irregular algorithms such as pointer chasing or recursive algorithms. The TASK directive takes effect only if you specify the -qsmp compiler option.
Syntax
TASK task_clause ,
block
END TASK
where task_clause is:
default_clause final_clause firstprivate_clause if_clause mergeable_clause private_clause shared_clause untied_clause
default_clause See “DEFAULT” on page 152. final_clause See “FINAL” on page 154.
Chapter 7. Parallel programming with XL Fortran 133 firstprivate_clause See “FIRSTPRIVATE” on page 155. if_clause See “IF” on page 156. mergeable_clause See “MERGEABLE” on page 159. private_clause See “PRIVATE” on page 161. shared_clause See “SHARED” on page 170. untied_clause See “UNTIED” on page 172.
Rules
A final task is a task that makes all its child tasks become final and included tasks. A final task is generated when either of the following conditions is true: v A FINAL clause is specified on a task construct and the FINAL clause expression evaluates to .TRUE.. v The generated task is a child task of a final task.
An undeferred task is a task whose execution is not deferred with respect to its generating task region. In other words, the generating task region is suspended until the undeferred task has finished running. An undeferred task is generated when an IF clause is specified on a task construct and the IF clause expression evaluates to .FALSE..
An included task is a task whose execution is sequentially included in the generating task region. In other words, an included task is undeferred and executed immediately by the encountering thread. An included task is generated when the generated task is a child task of a final task.
A merged task is a task that has the same data environment as that of its generating task region. A merged task might be generated when both the following conditions are true: v A MERGEABLE clause is specified on a task construct. v The generated task is an undeferred task or an included task.
The following rules are true if no DEFAULT clause is specified with the enclosing TASK construct: v If the enclosing TASK construct is not lexically enclosed by a parallel region, dummy arguments that do not appear in any PRIVATE, FIRSTPRIVATE, LASTPRIVATE,orSHARED clause of the enclosing TASK construct are firstprivate. v A variable that is private in the innermost enclosing parallel construct is firstprivate in the TASK construct. v Local variables of a routine are firstprivate if there is no enclosing parallel construct. v A variable that is determined to be shared in all of the enclosing constructs, up to and including the innermost enclosing parallel construct, is shared.
134 XL Fortran: Optimization and Programming Guide The IF clause expression and the FINAL clause expression are evaluated outside of the task construct, and the evaluation order is not specified. Related reference: “FINAL” on page 154 “FIRSTPRIVATE” on page 155 “IF” on page 156 “MERGEABLE” on page 159 “DEFAULT” on page 152 “PRIVATE” on page 161 “SHARED” on page 170 “TASKWAIT” “UNTIED” on page 172 TASKWAIT Purpose
The TASKWAIT directive specifies a wait for child tasks to be completed that are generated by the current task.
Syntax
TASKWAIT
Related reference: “TASK / END TASK” on page 133 TASKYIELD Purpose
The TASKYIELD directive instructs the compiler that it can suspend the current task in favor of running a different task. The TASKYIELD region includes an explicit task scheduling point in the current task region.
Syntax
TASKYIELD
THREADLOCAL Purpose
You can use the THREADLOCAL directive to declare thread-specific common data. It is a possible method of ensuring that access to data that is contained within COMMON blocks is serialized.
In order to make use of this directive it is not necessary to specify the -qsmp compiler option, but the invocation command must be bgxlf_r, bgxlf90_r, bgxlf95_r, bgxlf2003_r,orbgxlf2008_r to link the necessary libraries.
Chapter 7. Parallel programming with XL Fortran 135 Syntax
,
THREADLOCAL / common_block_name / ::
Rules
You can only declare named blocks as THREADLOCAL. All rules and constraints that normally apply to named common blocks apply to common blocks that are declared as THREADLOCAL. See the COMMON statement in the XL Fortran Language Reference for more information on the rules and constraints that apply to named common blocks.
The THREADLOCAL directive must appear in the specification_part of the scoping unit. If a common block appears in a THREADLOCAL directive, it must also be declared within a COMMON statement in the same scoping unit. The THREADLOCAL directive may occur before or after the COMMON statement. See Main program in the XL Fortran Language Reference for more information on the specification_part of the scoping unit.
A common block cannot be given the THREADLOCAL attribute if it is declared within a PURE subprogram.
Members of a THREADLOCAL common block must not appear in NAMELIST statements.
A common block that is use-associated must not be declared as THREADLOCAL in the scoping unit that contains the USE statement.
Any pointers declared in a THREADLOCAL common block are not affected by the -qinit=f90ptr compiler option.
Objects within THREADLOCAL common blocks may be used in parallel loops and parallel sections. However, these objects are implicitly shared across the iterations of the loop, and across code blocks within parallel sections. In other words, within a scoping unit, all accessible common blocks, whether declared as THREADLOCAL or not, have the SHARED attribute within parallel loops and sections in that scoping unit.
If a common block is declared as THREADLOCAL within a scoping unit, any subprogram that declares or references the common block, and that is directly or indirectly referenced by the scoping unit, must be executed by the same thread executing the scoping unit. If two procedures that declare common blocks are executed by different threads, then they would obtain different copies of the common block, provided that the common block had been declared THREADLOCAL. Threads can be created in one of the following ways: v Explicitly, via pthreads library calls v Implicitly by the compiler for parallel loop execution v Implicitly by the compiler for parallel section execution.
136 XL Fortran: Optimization and Programming Guide If a common block is declared to be THREADLOCAL in one scoping unit, it must be declared to be THREADLOCAL in every scoping unit that declares the common block.
If a THREADLOCAL common block that does not have the SAVE attribute is declared within a subprogram, the members of the block become undefined at subprogram RETURN or END, unless there is at least one other scoping unit in which the common block is accessible that is making a direct or indirect reference to the subprogram.
You cannot specify the same common_block_name for both a THREADLOCAL directive and a THREADPRIVATE directive.
Example 1: The following procedure "FORT_SUB" is invoked by two threads: SUBROUTINE FORT_SUB(IARG) INTEGER IARG
CALL LIBRARY_ROUTINE1() CALL LIBRARY_ROUTINE2() ... END SUBROUTINE FORT_SUB SUBROUTINE LIBRARY_ROUTINE1() COMMON /BLOCK/ R ! The SAVE attribute is required for the SAVE /BLOCK/ ! common block because the program requires ! that the block remain defined after !IBM* THREADLOCAL /BLOCK/ ! library_routine1 is invoked. R = 1.0 ... END SUBROUTINE LIBRARY_ROUTINE1 SUBROUTINE LIBRARY_ROUTINE2() COMMON /BLOCK/ R SAVE /BLOCK/ !IBM* THREADLOCAL /BLOCK/
...=R ... END SUBROUTINE LIBRARY_ROUTINE2
Example 2: "FORT_SUB" is invoked by multiple threads. This is an invalid example because "FORT_SUB" and "ANOTHER_SUB" both declare /BLOCK/ to be THREADLOCAL. They intend to share the common block, but they are executed by different threads. SUBROUTINE FORT_SUB() COMMON /BLOCK/ J INTEGER :: J !IBM* THREADLOCAL /BLOCK/ ! Each thread executing FORT_SUB ! obtains its own copy of /BLOCK/ INTEGER A(10)
... !IBM* INDEPENDENT DO INDEX = 1,10 CALL ANOTHER_SUB(A(I)) END DO ...
END SUBROUTINE FORT_SUB SUBROUTINE ANOTHER_SUB(AA) ! Multiple threads are used to execute ANOTHER_SUB INTEGER AA COMMON /BLOCK/ J ! Each thread obtains a new copy of the INTEGER :: J ! common block /BLOCK/
Chapter 7. Parallel programming with XL Fortran 137 !IBM* THREADLOCAL /BLOCK/ ... AA = J ! The value of ’J’ is undefined. END SUBROUTINE ANOTHER_SUB
One or more sample programs in the directory /opt/ibmcmp/xlf/bg/14.1/samples/ modules/threadlocal illustrate how to use the THREADLOCAL directive and create threads in C. Related reference:
See -qdirective in the Compiler Reference
See -qinit in the Compiler Reference
See COMMON in the Language Reference
See Main program in the Language Reference THREADPRIVATE Purpose
The THREADPRIVATE directive allows you to specify named common blocks and named variables as private to a thread but global within that thread. Once you declare a common block or variable THREADPRIVATE, each thread in the team maintains a separate copy of that common block or variable. Data written to a THREADPRIVATE common block or variable remains private to that thread and is not visible to other threads in the team.
In the serial and MASTER sections of a program, only the master thread's copy of the named common block and variable is accessible.
Use the COPYIN clause on the PARALLEL, PARALLEL DO, PARALLEL SECTIONS or PARALLEL WORKSHARE directives to specify that upon entry into a parallel region, data in the master thread's copy of a named common block or named variable is copied to each thread's private copy of that common block or variable.
The THREADPRIVATE directive only takes effect if you specify the -qsmp compiler option.
Syntax
THREADPRIVATE ( threadprivate_entity_list )
where threadprivate_entity_list is:
variable_name / common_block_name /
common_block_name is the name of a common block to be made private to a thread.
138 XL Fortran: Optimization and Programming Guide variable_name is the name of a variable to be made private to a thread.
Rules
You cannot specify a THREADPRIVATE variable, common block, or the variables that comprise that common block in a PRIVATE, FIRSTPRIVATE, LASTPRIVATE, SHARED,orREDUCTION clause.
A THREADPRIVATE variable must have the SAVE attribute. For variables or common blocks declared in the scope of a module, the SAVE attribute is implied. If you declare the variable outside of the scope of the module, the SAVE attribute must be specified.
In THREADPRIVATE directives, you can only specify named variables and named common blocks.
A variable can only appear in a THREADPRIVATE directive in the scope in which it is declared, and a THREADPRIVATE variable or common block may only appear once in a given scope. The variable must not be an element of a common block, or be declared in an EQUIVALENCE statement.
You cannot specify the same common_block_name for both a THREADPRIVATE directive and a THREADLOCAL directive.
All rules and constraints that apply to named common blocks also apply to common blocks declared as THREADPRIVATE. See the COMMON statement in the XL Fortran Language Reference.
If you declare a common block as THREADPRIVATE in one scoping unit, you must declare it as THREADPRIVATE in all other scoping units in which it is declared.
On entry into any parallel region, a THREADPRIVATE variable, or a variable in a THREADPRIVATE common block specified in a COPYIN clause is subject to the criteria stated in the Rules section for the COPYIN clause.
On entry into the first parallel region of the program, THREADPRIVATE variables or variables within a THREADPRIVATE common block not specified in a COPYIN clause are subject to the following criteria: v If the variable has the ALLOCATABLE attribute, the initial allocation status of each copy of that variable is not currently allocated. v If the variable has the POINTER attribute, and that pointer is disassociated through either explicit or default initialization, the association status of each copy of that variable is disassociated. Otherwise, the association status of the pointer is undefined. v If the variable has neither the ALLOCATABLE nor the POINTER attribute and is defined through either explicit or default initialization, then each copy of that variable is defined. If the variable is undefined, then each copy of that variable is undefined.
On entry into subsequent parallel regions of the program, THREADPRIVATE variables, or variables within a THREADPRIVATE common block not specified in a COPYIN clause, are subject to the following criteria:
Chapter 7. Parallel programming with XL Fortran 139 v If you are using the OMP_DYNAMIC environment variable, or the omp_set_dynamic subroutine to enable dynamic threads and: – If the number of threads is smaller than the number of threads in the previous region, and if a THREADPRIVATE object is referenced in both regions, then threads with the same thread number in their respective regions will reference the same copy of that variable. – If the number of threads is larger than the number of threads in the previous region, then the definition and association status of a THREADPRIVATE object is undefined, and the allocation status is undefined. v If dynamic threads are disabled, the definition, association, or allocation status and definition, if the thread's copy of the variable was defined, is retained.
You cannot access the name of a common block by use association or host association. Thus, a named common block can only appear on a THREADPRIVATE directive if the common block is declared in the scoping unit that contains the THREADPRIVATE directive. However, you can access the variables in the common block by use association or host association. For more information, see Host and Use association in the XL Fortran Language Reference.
The -qinit=f90ptr compiler option does not affect pointers that you have declared in a THREADPRIVATE common block.
The DEFAULT clause does not affect variables in THREADPRIVATE common blocks.
Examples
Example 1: In this example, the PARALLEL DO directive invokes multiple threads that call SUB1. The common block BLK in SUB1 shares the data that is specific to the thread with subroutine SUB2, which is called by SUB1. PROGRAM TT INTEGER :: I, B(50)
!$OMP PARALLEL DO SCHEDULE(STATIC, 10) DO I=1, 50 CALL SUB1(I, B(I)) ! Multiple threads call SUB1. ENDDO END PROGRAM TT
SUBROUTINE SUB1(J, X) INTEGER :: J, X, A(100) COMMON /BLK/ A !$OMP THREADPRIVATE(/BLK/) ! Array a is private to each thread. ! ... CALL SUB2(J) X = A(J) + A(J + 50) ! ... END SUBROUTINE SUB1
SUBROUTINE SUB2(K) INTEGER :: C(100) COMMON /BLK/ C !$OMP THREADPRIVATE(/BLK/) ! ... C=K ! ... ! Since each thread has its own copy of ! common block BLK, the assignment of ! array C has no effect on the copies of ! that block owned by other threads. END SUBROUTINE SUB2
140 XL Fortran: Optimization and Programming Guide Example 2: In this example, each thread has its own copy of the common block ARR in the parallel section. If one thread initializes the common block variable TEMP, the initial value is not visible to other threads. PROGRAM ABC INTEGER :: I, TEMP(100), ARR1(50), ARR2(50) COMMON /ARR/ TEMP !$OMP THREADPRIVATE(/ARR/) INTERFACE SUBROUTINE SUBS(X) INTEGER :: X(:) END SUBROUTINE END INTERFACE ! ... !$OMP PARALLEL SECTIONS !$OMP SECTION ! The thread has its own copy of the ! ... ! common block ARR. TEMP(1:100:2) = -1 TEMP(2:100:2) = 2 CALL SUBS(ARR1) ! ... !$OMP SECTION ! The thread has its own copy of the ! ... ! common block ARR. TEMP(1:100:2) = 1 TEMP(2:100:2) = -2 CALL SUBS(ARR2) ! ... !$OMP END PARALLEL SECTIONS ! ... PRINT *, SUM(ARR1), SUM(ARR2) END PROGRAM ABC
SUBROUTINE SUBS(X) INTEGER :: K, X(:), TEMP(100) COMMON /ARR/ TEMP !$OMP THREADPRIVATE(/ARR/) ! ... DOK=1,UBOUND(X, 1) X(K) = TEMP(K) + TEMP(K + 1) ! The thread is accessing its ! own copy of ! the common block. ENDDO ! ... END SUBROUTINE SUBS
The expected output for this program is: 50 -50
Example 3: In the following example, local variables outside of a common block are declared THREADPRIVATE. MODULE MDL INTEGER :: A(2) INTEGER, POINTER :: P INTEGER, TARGET :: T !$OMP THREADPRIVATE(A, P) END MODULE MDL
PROGRAM MVAR USE OMP_LIB USE MDL
INTEGER :: I
CALL OMP_SET_NUM_THREADS(2)
Chapter 7. Parallel programming with XL Fortran 141 A = (/1, 2/) T=4 P=>T
!$OMP PARALLEL PRIVATE(I) COPYIN(A, P) I = OMP_GET_THREAD_NUM() IF (I .EQ. 0) THEN A(1) = 100 T=5 ELSE IF (I .EQ. 1) THEN A(2) = 200 END IF !$OMP END PARALLEL
!$OMP PARALLEL PRIVATE(I) I = OMP_GET_THREAD_NUM() IF (I .EQ. 0) THEN PRINT *, ’A(2) = ’, A(2) ELSE IF (I .EQ. 1) THEN PRINT *, ’A(1) = ’, A(1) PRINT *, ’P => ’, P END IF !$OMP END PARALLEL
END PROGRAM MVAR
If dynamic threads mechanism is disabled, the expected output is: A(2) = 2 A(1) = 1 P=>5 or A(1) = 1 P=>5 A(2) = 2 Related reference:
See COMMON in the Language Reference “OMP_DYNAMIC” on page 79 “omp_set_dynamic(enable_expr)” on page 187 “PARALLEL / END PARALLEL” on page 111 “PARALLEL DO / END PARALLEL DO” on page 113 “PARALLEL SECTIONS / END PARALLEL SECTIONS” on page 117 TM_ATOMIC / END TM_ATOMIC
The TM_ATOMIC directive indicates a transactional atomic region.
Syntax
!TM$ TM_ATOMIC safe_mode
!TM$ END TM_ATOMIC
safe_mode Using the safe_mode clause reduces overhead and increases performance.
142 XL Fortran: Optimization and Programming Guide However, if safe_mode is specified, irrevocable actions are not checked at runtime. The run result is undefined if an irrevocable action occurs during the execution.
Usage
The transactional memory directive is enabled with the -qtm compiler option. To compile your program with transactional memory, you must use thread safe invocation commands.
This directive must be placed immediately before the code block of the transactional atomic region.
A transactional atomic region must be one of the following structured blocks: v A block of executable statements with a single entry at the top and a single exit at the bottom v An OpenMP construct
You cannot branch into or break out from the middle of a transactional atomic region. For example, you cannot use a GO TO statement to transfer control out of a transactional atomic region.
When you use the END=, ERR=,orEOR= I/O statement specifier for branching, specify a statement that is within the same transactional atomic region as the I/O statement.
Transactional atomic regions can be nested and the maximum nesting level is 222. However, the atomicity, consistency, and isolation properties are provided at the outermost transactional atomic region. v Stopping an inner transaction causes the corresponding outer transaction to stop. v An inner transaction does not commit data at the end of its transactional atomic region. Instead, the data of the inner transaction is committed later when the data of the corresponding outer transaction is committed. v If a conflict occurs in a nested transaction, the thread is rolled back to the beginning of the outermost transaction.
Example USE OMP_LIB
INTEGER :: i INTEGER :: data_arr(100)
!$OMP PARALLEL DO DOi=1,100
! Use the TM_ATOMIC directive to indicate a transactional atomic region. !TM$ TM_ATOMIC data_arr(i)=i+1 !TM$ END TM_ATOMIC
END DO !$OMP END PARALLEL DO
Related information v The "-qtm" compiler option v Execution modes
Chapter 7. Parallel programming with XL Fortran 143 v Routines for transactional memory v Environment variables for transactional memory v "STOP" WORKSHARE / END WORKSHARE Purpose
The WORKSHARE directive allows you to parallelize the execution of array operations. A WORKSHARE directive divides the tasks associated with an enclosed block of code into units of work. When a team of threads encounters a WORKSHARE directive, the threads in the team share the tasks, so that each unit of work executes exactly once.
The WORKSHARE directive only takes effect if you specify the -qsmp compiler option.
Syntax
WORKSHARE
block
END WORKSHARE NOWAIT
block is a structured block of statements that allows work sharing within the lexical extent of the WORKSHARE construct. The execution of statements are synchronized so that statements whose result is a dependent on another statement are evaluated before that result is required. The block can contain any of the following: v Array assignment statements v ATOMIC directives v CRITICAL constructs v FORALL constructs v FORALL statements v PARALLEL construct v PARALLEL DO construct v PARALLEL SECTION construct v PARALLEL WORKSHARE construct v Scalar assignment statements v WHERE constructs v WHERE statements The transformational intrinsic functions you can use as part of an array operation are:
144 XL Fortran: Optimization and Programming Guide v ALL v MATMUL v PRODUCT v ANY v MAXLOC v RESHAPE v COUNT v MAXVAL v SPREAD v CSHIFT v MINLOC v SUM v DOT_PRODUCT v MINVAL v TRANSPOSE v EOSHIFT v PACK v UNPACK
The block can also contain statements bound to lexically enclosed PARALLEL constructs. These statements are not restricted. Any user–defined function calls within the block must be elemental.
Statements enclosed in a WORKSHARE directive are divided into units of work. The definition of a unit of work varies according to the statement evaluated. A unit of work is defined as follows: v Array expressions: Evaluation of each element of an array expression is a unit of work. Any of the transformational intrinsic functions listed above may be divided into any number of units of work. v Assignment statements: In an array assignment statement, the assignment of each element in the array is a unit of work. For scalar assignment statements, the assignment operation is a unit of work. v Constructs: Evaluation of each CRITICAL construct is a unit of work. Each PARALLEL construct contained within a WORKSHARE construct is a single unit of work. New teams of threads execute the statements contained within the lexical extent of the enclosed PARALLEL constructs. In FORALL constructs or statements, the evaluation of the mask expression, expressions occurring in the specification of the iteration space, and the masked assignments are units of work. In WHERE constructs or statements, the evaluation of the mask expression and the masked assignments are units of work. v Directives: The update of each scalar variable for an ATOMIC directive and its assignments is a unit of work. v ELEMENTAL functions: If the argument to an ELEMENTAL function is an array, then the application of the function to each element of an array is a unit of work.
If none of the above definitions apply to a statement within the block, then that statement is a unit of work.
Rules
In order to ensure that the statements within a WORKSHARE construct execute in parallel, the construct must be enclosed within a parallel region. Threads encountering a WORKSHARE construct outside the dynamic extent of a parallel region will evaluate the statements within the construct serially.
A WORKSHARE directive binds to the closest enclosing PARALLEL region if one exists.
You must not nest work-sharing regions that bind to the same PARALLEL region.
You must not specify a WORKSHARE directive within the CRITICAL, MASTER, or ORDERED regions.
You must not specify BARRIER, MASTER,orORDERED directives within a WORKSHARE region.
Chapter 7. Parallel programming with XL Fortran 145 If an array assignment, scalar assignment, a masked array assignment or a FORALL assignment assigns to a private variable in the block, the result is undefined.
If an array expression in the block references the value, association status or allocation status of private variables, the value of the expression is undefined unless each thread computes the same value.
If you do not specify a NO WAIT clause at the end of a WORKSHARE construct, a BARRIER directive is implied.
A WORKSHARE construct must be encountered by all threads in the team or by none at all.
Examples
Example 1: In the following example, the WORKSHARE directive evaluates the masked expressions in parallel. !$OMP WORKSHARE FORALL (I=1:N,AA(1, I) == 0) AA(1, I) = I BB = TRANSPOSE(AA) CC = MATMUL(AA, BB) !$OMP ATOMIC S=S+SUM(CC) !$OMP END WORKSHARE
Example 2: The following example includes a user defined ELEMENTAL as part of a WORKSHARE construct. !$OMP WORKSHARE WHERE (AA(1, :) /= 0.0) AA(1, :)=1/AA(1, :) DD = TRANS(AA(1, :)) !$OMP END WORKSHARE
ELEMENTAL REAL FUNCTION TRANS(ELM) RESULT(RES) REAL, INTENT(IN) :: ELM RES=ELM*ELM+4 END FUNCTION Related reference: “ATOMIC” on page 93 “BARRIER” on page 97 “CRITICAL / END CRITICAL” on page 99 “PARALLEL WORKSHARE / END PARALLEL WORKSHARE” on page 119
See -qsmp in the Compiler Reference Directive clauses You can use directive clauses to specify additional information to directives. Global rules for directive clauses
You must not specify a variable or common block name more than once in a clause.
A variable, common block name, or variable name that is a member of a common block must not appear in more than one clause on the same directive, with the following exceptions:
146 XL Fortran: Optimization and Programming Guide v You can define a named common block or named variable as FIRSTPRIVATE and LASTPRIVATE for the same directive. v A variable appearing in a NUM_THREADS clause can appear in another clause for the same directive. v A variable appearing in a IF clause can appear in another clause for the same directive.
If you do not specify a clause that changes the scope of a variable, the default scope for variables affected by a directive is SHARED.
A local variable with the SAVE or STATIC attribute declared in a procedure referenced a parallel region has an implicit SHARED attribute. A local variable without the SAVE or STATIC attribute declared in a procedure referenced a parallel region has an implicit PRIVATE attribute.
Members of common blocks and variables of modules declared in a procedure referenced within the dynamic extent of a parallel region have an implicit SHARED attribute, unless they are THREADLOCAL or THREADPRIVATE common blocks and module variables.
While a parallel or work-sharing construct is running, a variable or variable subobject used in a PRIVATE, FIRSTPRIVATE, LASTPRIVATE or REDUCTION clause of the directive must not be referenced, become defined, become undefined, have its association status or allocation status changed, or appear as an actual argument: v In a scoping unit other than the one in which the directive construct appears v In a variable format expression
You can declare a variable as PRIVATE, FIRSTPRIVATE, LASTPRIVATE,or REDUCTION, even if that variable is already storage associated with other variables. Storage association may exist for variables declared in EQUIVALENCE statements or in COMMON blocks. If a variable is storage associated with a PRIVATE, FIRSTPRIVATE, LASTPRIVATE,orREDUCTION variable, then: v The contents, allocation status and association status of the variable that is storage associated with the PRIVATE, FIRSTPRIVATE, LASTPRIVATE or REDUCTION variable are undefined on entry to the parallel construct. v The allocation status, association status and the contents of the associated variable become undefined if you define the PRIVATE, FIRSTPRIVATE, LASTPRIVATE or REDUCTION variable or if you define that variable's allocation or association status. v The allocation status, association status and the contents of the PRIVATE, FIRSTPRIVATE, LASTPRIVATE or REDUCTION variable become undefined if you define the associated variable or if you define the associated variable's allocation or association status. Pointers and OpenMP API
OpenMP API allows a variable or variable subobject of a PRIVATE clause to have the POINTER or ALLOCATABLE attribute. The association status of the pointer is undefined at thread creation and when the thread is destroyed.
See the following topics for more information about the directive clauses:
Chapter 7. Parallel programming with XL Fortran 147 COLLAPSE FIRSTPRIVATE PRIVATE COPYIN LASTPRIVATE REDUCTION COPYPRIVATE NUM_THREADS SCHEDULE DEFAULT ORDERED SHARED IF UNTIED
COLLAPSE Purpose
Specifying the COLLAPSE clause allows you to parallelize multiple loops in a nest without introducing nested parallelism.
Syntax
COLLAPSE ( n )
n is a positive constant integer expression
Rules v Only one collapse clause is allowed on a worksharing DO or PARALLEL DO directive v The specified number of loops must be present lexically. That is, none of the loops can be in a called subroutine. v The loops must form a rectangular iteration space and the bounds and stride of each loop must be invariant over all the loops. v If the loop indices are of different size, the index with the largest size will be used for the collapsed loop. v The loops must be perfectly nested; that is, there is no intervening code nor any OpenMP directive between the loops which are collapsed. v The associated do-loops must be structured blocks. Their execution must not be terminated by an EXIT statement. v If multiple loops are associated to the loop construct, only an iteration of the innermost associated loop may be curtailed by a CYCLE statement. If multiple loops are associated to the loop construct, there must be no branches to any of the loop termination statements except for the innermost associated loop. Ordered construct During execution of an iteration of a loop or a loop nested within a loop region, the executing thread must not execute more than one ordered region which binds to the same loop region. As a consequence, if multiple loops are associated to the loop construct by a collapse clause, the ordered construct has to be located inside all associated loops. LASTPRIVATE clause When a LASTPRIVATE clause appears on the directive that identifies a work-sharing construct, the value of each new list item from the sequentially last iteration of the associated loops is assigned to the original list item even if a collapse clause is associated with the loop Other SMP and performance directives The STREAM_UNROLL, UNROLL, UNROLL_AND_FUSE, and NOUNROLL_AND_FUSE directives cannot be used for any of the loops
148 XL Fortran: Optimization and Programming Guide associated with the COLLAPSE clause loop nest. The INDEPENDENT directive can be used for any of the loops associated with the COLLAPSE clause.
Examples
In Example 1 and Example 2 the loops over k and j are collapsed and their iteration space is executed by all threads of the current team.
Example 1 !$omp do collapse(2) private(i,j,k) do k = kl, ku, ks do j = jl, ju, js do i = il, iu, is call bar(a,i,j,k) enddo enddo enddo !$omp end do
Example 2 program test !$omp parallel !$omp do private(j,k) collapse(2) lastprivate(jlast, klast) do k = 1,2 do j = 1,3 jlast=j klast=k enddo enddo !$omp end do !$omp single print *, klast, jlast !$omp end single !$omp end parallel end program test
Output: 23
Example 3
As both loops are collapsed into one, the ordered construct has to be inside all loops associated to the for construct. As an iteration may not execute more than one ordered region, this program would be incorrect without the collapse(2) clause. program test !$omp parallel num_threads(2) !$omp do collapse(2) ordered private(j,k) schedule(static,3) do k = 1,3 do j = 1,2 !$omp ordered print *, k, j !$omp end ordered enddo enddo !$omp end do !$omp end parallel end program test
Output:
Chapter 7. Parallel programming with XL Fortran 149 11 12 21 22 31 32 Related reference: ORDERED / END ORDERED DO / END DO PARALLEL DO / END PARALLEL DO COPYIN Purpose
If you specify the COPYIN clause, the master thread's copy of each variable, or common block declared in the copyin_entity_list is duplicated at the beginning of a parallel region. Each thread in the team that will execute within that parallel region receives a private copy of all entities in the copyin_entity_list. All variables declared in the copyin_entity_list must be THREADPRIVATE or members of a common block that appears in a THREADPRIVATE directive.
Syntax
COPYIN ( copyin_entity_list )
copyin_entity
variable_name / common_block_name /
variable is a THREADPRIVATE variable, or THREADPRIVATE variable in a common block. common_block_name is a THREADPRIVATE common block name.
Rules
If you specify a COPYIN clause, you cannot: v specify the same entity name more than once in a copyin_entity_list. v specify the same entity name in separate COPYIN clauses on the same directive. v specify both a common block name and any variable within that same named common block in a copyin_entity_list. v specify both a common block name and any variable within that same named common block in different COPYIN clauses on the same directive. v specify a variable that contains ALLOCATABLE components.
150 XL Fortran: Optimization and Programming Guide When the master thread of a team of threads reaches a directive containing the COPYIN clause, thread's private copy of a variable or common block specified in the COPYIN clause will have the same value as the master thread's copy.
On entry into any parallel region, a THREADPRIVATE variable, or a variable in a THREADPRIVATE common block is subject to the following criteria when declared in a COPYIN clause: v If the variable has the POINTER attribute and the master thread's copy of the variable is associated with a target, then each copy of that variable is associated with the same target. If the master thread's pointer is disassociated, then each copy of that variable is disassociated. If the master thread's copy of the variable has an undefined association status, then each copy of that variable has an undefined association status. v Each copy of a variable without the POINTER attribute becomes defined with the value of the master thread's copy as if by intrinsic assignment.
If an allocatable array is specified in a COPYIN clause and it is allocated on entry into a parallel region, each thread copy of that array must be allocated with the same bounds and rank. Related reference: PARALLEL / END PARALLEL PARALLEL DO / END PARALLEL DO PARALLEL SECTIONS / END PARALLEL SECTIONS PARALLEL WORKSHARE / END PARALLEL WORKSHARE COPYPRIVATE Purpose
If you specify the COPYPRIVATE clause, the value of a private variable or pointer to a shared object from one thread in a team is copied into the corresponding variables of all other threads in that team. If the variable in copyprivate_entity_list is not a pointer, then the corresponding variables of all threads within that team are defined with the value of that variable. If the variable is a pointer, then the corresponding variables of all threads within that team are defined with the association status of the pointer. Integer pointers and assumed-size arrays must not appear in copyprivate_entity_list.
Syntax
COPYPRIVATE ( copyprivate_entity_list )
copyprivate_entity
variable / common_block_name /
Chapter 7. Parallel programming with XL Fortran 151 variable is a private variable within the enclosing parallel region common_block_name is a THREADPRIVATE common block name
Rules
If a common block is part of the copyprivate_entity_list, then it must appear in a THREADPRIVATE directive. Furthermore, the COPYPRIVATE clause treats a common block as if all variables within its object_list were specified in the copyprivate_entity_list.
A COPYPRIVATE clause must occur on an END SINGLE directive at the end of a SINGLE construct. The compiler evaluates a COPYPRIVATE clause before any threads have passed the implied BARRIER directive at the end of that construct. The variables you specify in copyprivate_entity_list must not appear in a PRIVATE or FIRSTPRIVATE clause for the SINGLE construct. If the END SINGLE directive occurs within the dynamic extent of a parallel region, the variables you specify in copyprivate_entity_list must be private within that parallel region.
A COPYPRIVATE clause must not appear on the same END SINGLE directive as a NOWAIT clause.
A THREADLOCAL common block, or members of that common block, are not permitted as part of a COPYPRIVATE clause.
If an allocatable array appears on a COPYPRIVATE clause, it must have an allocation status of allocated with the same bounds and rank in all threads that are affected by the COPYPRIVATE clause. Related reference: SINGLE / END SINGLE DEFAULT Purpose
If you specify the DEFAULT clause, all variables in the lexical extent of the parallel construct will have a scope attribute of default_scope_attr.
If you specify DEFAULT(NONE), there is no default scope attribute. Therefore, you must explicitly list each variable you use in the lexical extent of the parallel construct in a data scope attribute clause on the parallel construct, unless the variable is: v THREADPRIVATE v A member of a THREADPRIVATE common block. v A pointee v A loop iteration variable used only as a loop iteration variable for: – Sequential loops in the lexical extent of the parallel region, or, – Parallel do loops that bind to the parallel region v A variable that is only used in work-sharing constructs that bind to the parallel region, and is specified in a data scope attribute clause for each of the work-sharing constructs.
152 XL Fortran: Optimization and Programming Guide The DEFAULT clause specifies that all variables in the parallel construct share the same default scope attribute of either FIRSTPRIVATE, PRIVATE, SHARED,orno default scope attribute.
Syntax
DEFAULT ( default_scope_attr )
default_scope_attr is one of FIRSTPRIVATE, PRIVATE, SHARED,orNONE
Rules
If you specify DEFAULT(NONE) on a directive you must specify all named variables and all the leftmost names of referenced array sections, array elements, structure components, or substrings in the lexical extent of the directive construct in a FIRSTPRIVATE, LASTPRIVATE, PRIVATE, REDUCTION,orSHARED clause.
If you specify DEFAULT(FIRSTPRIVATE) on a directive, all named variables and all leftmost names of referenced array sections, array elements, structure components, or substrings in the lexical extent of the directive construct, including common block and use associated variables, but excluding POINTEEs and THREADLOCAL common blocks, have a FIRSTPRIVATE attribute to a thread as if they were listed explicitly in a FIRSTPRIVATE clause.
If you specify DEFAULT(PRIVATE) on a directive, all named variables and all leftmost names of referenced array sections, array elements, structure components, or substrings in the lexical extent of the directive construct, including common block and use associated variables, but excluding POINTEEs and THREADLOCAL common blocks, have a PRIVATE attribute to a thread as if they were listed explicitly in a PRIVATE clause.
If you specify DEFAULT(SHARED) on a directive, all named variables and all leftmost names of referenced array sections, array elements, structure components, or substrings in the lexical extent of the directive construct, excluding POINTEEs have a SHARED attribute to a thread as if they were listed explicitly in a SHARED clause.
The default behavior will be DEFAULT(SHARED) if you do not explicitly indicate a DEFAULT clause on a directive.
Example for thread-level speculative execution
The following example demonstrates the use of DEFAULT(PRIVATE) in SPECULATIVE directives. PROGRAM MAIN REAL(8) :: x REAL(8), DIMENSION(4) :: y
!SEP$ SPECULATIVE SECTIONS DEFAULT(PRIVATE), SHARED(y) x = 1.0
Chapter 7. Parallel programming with XL Fortran 153 y(1) = x !SEP$ SPECULATIVE SECTION x = 2.25 y(2) = x !SEP$ SPECULATIVE SECTION x = 3.5 y(3) = x !SEP$ SPECULATIVE SECTION x = 4.75 y(4) = x !SEP$ END SPECULATIVE SECTIONS
IF(ANY(y .NE. (/1.0, 2.25, 3.5, 4.75/))) STOP 11 END PROGRAM
Example for OpenMP
The following example demonstrates the use of DEFAULT(NONE) for OpenMP, and some of the rules for specifying the data scope attributes of variables in the parallel region. PROGRAM MAIN COMMON /COMBLK/ abc(10), def
! The loop iteration variable, i, is not required to be ! in data scope attribute clause. $OMP PARALLEL DEFAULT(NONE) SHARED(ABC)
! def is specified on the work-sharing DO, and is not required to be ! specified in a data scope attribute clause on the parallel region. !$OMP DO FIRSTPRIVATE(def) DO i = 1,10 ABC(i) = def END DO !$OMP END PARALLEL END PROGRAM Related reference: PARALLEL / END PARALLEL PARALLEL DO / END PARALLEL DO PARALLEL SECTIONS / END PARALLEL SECTIONS PARALLEL WORKSHARE / END PARALLEL WORKSHARE “TASK / END TASK” on page 133 FINAL Purpose
The FINAL clause is used with the TASK directive. If you specify a FINAL clause and the scalar_logical_expr evaluates to .TRUE., the generated task is a final task. All task constructs encountered inside a final task create final and included tasks.
Syntax
FINAL ( scalar_logical_expr )
Rules
You can specify only one FINAL clause on the TASK directive.
154 XL Fortran: Optimization and Programming Guide Related reference
“TASK / END TASK” on page 133 FIRSTPRIVATE Purpose
If you use the FIRSTPRIVATE clause, each thread has its own initialized local copy of the variables and common blocks in data_scope_entity_list.
The FIRSTPRIVATE clause can be specified for the same variables as the PRIVATE clause, and functions in a manner similar to the PRIVATE clause. The exception is the status of the variable upon entry into the directive construct; the FIRSTPRIVATE variable exists and is initialized for each thread entering the directive construct.
Syntax
FIRSTPRIVATE ( data_scope_entity_list )
Rules
A variable in a FIRSTPRIVATE clause must not be any of the following elements: v A pointee v An assumed-size array v A THREADLOCAL common block v A THREADPRIVATE common block or its members v A THREADPRIVATE variable v An allocatable scalar object
You cannot specify a variable in a FIRSTPRIVATE clause of a parallel construct if both the following conditions are true: v The variable appears in a namelist statement, variable format expression or in an expression for a statement function definition. v You reference the statement function, the variable format expression through formatted I/O, or the namelist through namelist I/O, within the parallel construct.
For a variable specified in the FIRSTPRIVATE clause, the status of the private copies is determined as follows: v If the variable has the POINTER attribute, the private copies of the FIRSTPRIVATE variable receive the same association status as the original copy as if by pointer assignment. v If the variable does not have the POINTER attribute, the initialization of the private copies occurs as if by intrinsic assignment. However, if the original variable is not currently allocated, the private copies have the same allocation status as the original copy.
If an allocatable array appears on a FIRSTPRIVATE clause, it must have an allocation status of allocated upon entrance into the parallel construct that contains the FIRSTPRIVATE clause.
Chapter 7. Parallel programming with XL Fortran 155 When individual members of a common block are privatized, the storage of the specified variable is no longer associated with the storage of the common block.
Any variable that is storage associated with a FIRSTPRIVATE variable is undefined on entrance into the parallel construct.
If a directive construct contains a FIRSTPRIVATE argument to a Message Passing Interface (MPI) routine performing non-blocking communication, the MPI communication must complete before the end of the construct.
Example for thread-level speculative execution
The following example demonstrates the use of FIRSTPRIVATE in SPECULATIVE directives. PROGRAM MAIN INTEGER(8) :: x, i8 REAL(8) :: rar(2), r8
r8 = 2.0_8 i8 = 9_8
!SEP$ SPECULATIVE SECTIONS FIRSTPRIVATE(r8, i8) rar(1) = r8 * 2.0 x=i8 !SEP$ SPECULATIVE SECTION rar(2) = r8 + i8 !SEP$ END SPECULATIVE SECTIONS END PROGRAM Related reference: DO / END DO PARALLEL / END PARALLEL PARALLEL DO / END PARALLEL DO PARALLEL SECTIONS / END PARALLEL SECTIONS PARALLEL WORKSHARE / END PARALLEL WORKSHARE SECTIONS / END SECTIONS SINGLE / END SINGLE “TASK / END TASK” on page 133 IF Purpose
If you specify the IF clause, the runtime environment evaluates whether the scalar_logical_expression is true or false. If scalar_logical_expression is: v true, the block is run in parallel. v false, the containing region is suspended and the generated task is immediately run as though it is in a distinct task region. Note that for the TASK directive, if the IF clause is evaluated to true, the block is not required to run in parallel.
156 XL Fortran: Optimization and Programming Guide Syntax
IF ( scalar_logical_expression )
Rules
The IF clause can be used in the PARALLEL, PARALLEL DO, PARALLEL SECTIONS, PARALLEL WORKSHARE, and TASK directives.
The IF clause may appear at most once in any directive.
By default, a nested parallel loop is serialized, regardless of the setting of the IF clause. You can change this default by using the -qsmp=nested_par compiler option.
An IF expression is evaluated outside of the context of the parallel construct. Any function reference in the IF expression must not have side effects. Related reference: “PARALLEL / END PARALLEL” on page 111 “PARALLEL DO / END PARALLEL DO” on page 113 “PARALLEL SECTIONS / END PARALLEL SECTIONS” on page 117 “PARALLEL WORKSHARE / END PARALLEL WORKSHARE” on page 119 “TASK / END TASK” on page 133 LASTPRIVATE Purpose
If you use the LASTPRIVATE clause, each variable and common block in data_scope_entity_list is PRIVATE, and the last value of each variable in data_scope_entity_list can be referred to outside of the construct of the directive. If you use the LASTPRIVATE clause with DO or PARALLEL DO, the last value is the value of the variable after the last sequential iteration of the loop. If you use the LASTPRIVATE clause with SECTIONS or PARALLEL SECTIONS, the last value is the value of the variable after the last SECTION of the construct. If the last iteration of the loop or last section of the construct does not define a LASTPRIVATE variable, the variable is undefined after the loop or construct.
The LASTPRIVATE clause functions in a manner similar to the PRIVATE clause and you should specify it for variables that match the same criteria. The exception is in the status of the variable on exit from the directive construct. The compiler determines the last value of the variable, and takes a copy of that value which it saves in the named variable for use after the construct. A LASTPRIVATE variable is undefined on entry to the construct if it is not a FIRSTPRIVATE variable.
Syntax
LASTPRIVATE ( data_scope_entity_list )
Chapter 7. Parallel programming with XL Fortran 157 Rules
A variable in a LASTPRIVATE clause must not be any of the following elements: v A pointee v An allocatable scalar object v An assumed-size array v A THREADLOCAL common block v A THREADPRIVATE common block or its members v A THREADPRIVATE variable
You cannot specify a variable in a LASTPRIVATE clause of a parallel construct if both the following conditions are true: v The variable appears in a namelist statement, variable format expression or in an expression for a statement function definition. v You reference the statement function, the variable format expression through formatted I/O, or the namelist through namelist I/O, within the parallel construct.
A LASTPRIVATE variable must be definable.
For a variable specified in a LASTPRIVATE clause, v If the variable has the POINTER attribute, the original variable is updated as if by pointer assignment. v If the variable does not have the POINTER attribute, the original variable is updated as if by intrinsic assignment.
If an allocatable array appears on a LASTPRIVATE clause, its allocation status must be allocated when it enters into the parallel construct that contains the LASTPRIVATE clause. The private copies of the LASTPRIVATE variable in the sequentially last iteration or lexically last section must have an allocation status of allocated. They must have the same bounds and rank as the corresponding LASTPRIVATE variable when they exit from that iteration or section.
When individual members of a common block are privatized, the storage of the specified variable is no longer associated with the storage of the common block.
Any variable that is storage associated with a LASTPRIVATE variable is undefined on entrance into the parallel construct.
If you specify a variable as LASTPRIVATE on a work-sharing directive, and you have specified a NOWAIT clause on that directive, you cannot use that variable between the end of the work-sharing construct and a BARRIER directive.
Variables that you specify as LASTPRIVATE to a parallel construct become defined at the end of the construct. If you have concurrent definitions or uses of LASTPRIVATE variables on multiple threads, you must ensure that the threads are synchronized at the end of the construct when the variables become defined. For example, if multiple threads encounter a PARALLEL construct with a LASTPRIVATE variable, you must synchronize the threads when they reach the END PARALLEL directive, because the LASTPRIVATE variable becomes defined at END PARALLEL. Therefore the whole PARALLEL construct must be enclosed within a synchronization construct.
158 XL Fortran: Optimization and Programming Guide If a directive construct contains a LASTPRIVATE argument to a Message Passing Interface (MPI) routine performing non-blocking communication, the MPI communication must complete before the end of that construct.
Example for thread-level speculative execution
The following example demonstrates the use of LASTPRIVATE in SPECULATIVE directives. PROGRAM MAIN INTEGER :: flag1, flag2 INTEGER, ALLOCATABLE :: i2a(:) INTEGER :: n
flag1 = 0 flag2 = 0 ALLOCATE(i2a(2)) i2a = -99 n=80
!SEP$ SPECULATIVE SECTIONS LASTPRIVATE(i2a) IF(i2a(1) == -99) flag1 = 0 i2a(1) = 8 !SEP$ SPECULATIVE SECTION IF(i2a(2) == -99) flag2 = 0 i2a(2) = n i2a(2) = i2a(2) * 10 !SEP$ END SPECULATIVE SECTIONS IF(flag1 == 1) STOP 21 IF(flag2 == 1) STOP 22 END PROGRAM
Example for OpenMP
The following example shows the proper use of a LASTPRIVATE variable after a NOWAIT clause. !$OMP PARALLEL !$OMP DO LASTPRIVATE(k) DO i = 1,10 k=i+1 END DO
!$OMP END DO NOWAIT k = ... **ERROR** ! The reference to k must occur after a barrier. !$OMP BARRIER k = ... ! this reference to k is valid. !$OMP END PARALLEL END Related reference: DO / END DO PARALLEL DO / END PARALLEL DO PARALLEL SECTIONS / END PARALLEL SECTIONS SECTIONS / END SECTIONS MERGEABLE Purpose
The MERGEABLE clause is used with the TASK directive. If you specify a MERGEABLE clause and the generated task is an undeferred task or an included task, a merged task might be generated.
Chapter 7. Parallel programming with XL Fortran 159 Syntax
MERGEABLE
Related reference
“TASK / END TASK” on page 133 NUM_THREADS Purpose
The NUM_THREADS clause allows you to specify the number of threads used in a parallel region. Subsequent parallel regions are not affected. The NUM_THREADS clause takes precedence over the number of threads specified using the omp_set_num_threads library routine or the environment variable OMP_NUM_THREADS.
Syntax
NUM_THREADS ( scalar_integer_expression )
Rules
The value of scalar_integer_expression must be a positive. Evaluation of the expression occurs outside the context of the parallel region. Any function calls that appear in the expression and change the value of a variable referenced in the expression will have unspecified results.
If you are using the environment variable OMP_DYNAMIC to enable dynamic threads, scalar_integer_expression defines the maximum number of threads available in the parallel region.
You must specify the omp_set_nested library routine or set the OMP_NESTED environment variable when including the NUM_THREADS clause as part of a nested parallel regions otherwise, the execution of that parallel region is serialized. Related reference: PARALLEL / END PARALLEL PARALLEL DO / END PARALLEL DO PARALLEL SECTIONS / END PARALLEL SECTIONS PARALLEL WORKSHARE / END PARALLEL WORKSHARE ORDERED Purpose
Specifying the ORDERED clause on a work–sharing construct allows you to specify the ORDERED directive within the dynamic extent of a parallel loop.
160 XL Fortran: Optimization and Programming Guide Syntax
ORDERED
Rules
The ORDERED clause applies to the following directives: Related reference: “DO / END DO” on page 100 “PARALLEL DO / END PARALLEL DO” on page 113 PRIVATE Purpose
If you specify the PRIVATE clause on one of the directives listed below, each thread in a team has its own uninitialized local copy of the variables and common blocks in data_scope_entity_list.
You should specify a variable in the PRIVATE clause if its value is calculated by a single thread and that value is not dependent on any other thread, if it is defined before it is used in the construct, and if its value is not used after the construct ends. Copies of the PRIVATE variable exist, locally, on each thread. Each thread receives its own uninitialized copy of the PRIVATE variable. All thread variables within the lexical extent of the directive construct have the PRIVATE attribute by default.
Syntax
PRIVATE ( data_scope_entity_list )
Rules
A variable in the PRIVATE clause must not be any of the following elements: v A pointee v An assumed-size array v A THREADLOCAL common block v A THREADPRIVATE common block or its members v A THREADPRIVATE variable or the variable equivalenced with a THREADPRIVATE variable
You cannot specify a variable in a PRIVATE clause of a parallel construct if: v the variable appears in a namelist statement, variable format expression or in an expression for a statement function definition, and, v you reference the statement function, the variable format expression through formatted I/O, or the namelist through namelist I/O, within the parallel construct.
Chapter 7. Parallel programming with XL Fortran 161 When individual members of a common block are privatized, the storage of the specified variable is no longer associated with the storage of the common block.
A variable that appears in the REDUCTION clause of a parallel construct can also appear in a PRIVATE clause on a work-sharing construct.
If a directive construct contains a PRIVATE argument to a Message Passing Interface (MPI) routine performing non-blocking communication, the MPI communication must complete before the end of that construct.
A variable name in the data_scope_entity_list of the PRIVATE clause can be an allocatable array. If the allocatable array is allocated on entry to a parallel region, the private copies of the array has an allocation status of allocated and has the same rank and bounds as the PRIVATE variable. If the allocatable array is unallocated on entry to a parallel region, the private copies of the array has an allocation status of unallocated.
Local variables without the SAVE or STATIC attributes in referenced subprograms in the dynamic extent of a directive construct have an implicit PRIVATE attribute.
Example for thread-level speculative execution
The following example demonstrates the use of PRIVATE in SPECULATIVE directives. PROGRAM MAIN INTEGER, TARGET :: e INTEGER, POINTER :: p INTEGER :: f(10), flag
flag = 0 e=-1 p=>e
!SEP$ SPECULATIVE DO PRIVATE(e) DOi=1,10 e=i f(i) = e IF(p /= -1) flag = 1 END DO !SEP$ END SPECULATIVE DO
IF(flag == 1) STOP 11 END PROGRAM
Examples for OpenMP
Example 1: The following example demonstrates the proper use of a PRIVATE variable that is used to define a statement function. A commented line shows the invalid use. Since J appears in a statement function, the statement function cannot be referenced within the parallel construct for which J is PRIVATE. INTEGER :: arr(10),j=17 ISTFNC() = j
!$OMP PARALLEL DO PRIVATE(j) DOi=1,10 j=i ! arr(i) = ISTFNC() **ERROR** A reference to ISTFNC would ! make the PRIVATE(J) clause invalid.
162 XL Fortran: Optimization and Programming Guide ARR(i) = j END DO PRINT *, arr END
Example 2: The following example demonstrates the use of allocatable arrays on a PRIVATE clause: USE OMP_LIB REAL, ALLOCATABLE :: temp(:,:) REAL :: arr(4, 20, 20) INTEGER :: thd
ALLOCATE(temp(20, 20)) !$OMP PARALLEL PRIVATE(thd, temp) NUM_THREADS(4)
! Private copies of "temp" are allocated with the same ! bounds and shape of the original "temp". thd = OMP_GET_THREAD_NUM() IF(MOD(thd, 2) .EQ. 0) THEN temp = RESHAPE((/(i, i=1, 400)/), (/20, 20/)) ELSE temp = RESHAPE((/(i, i=1, 800, 2)/), (/20, 20/)) ENDIF arr(thd + 1, :, :) = temp
! Private copies of "temp" are deallocated. !$OMP END PARALLEL DEALLOCATE(temp) END
Note: If the machine has less than 4 CPUs, you must set OMP_THREAD_LIMIT=4.
Example 3: The following example demonstrates the persistence of the original value of the PRIVATE variables after exit from a parallel region: PROGRAM MAIN INTEGER :: i, j
i=1 j=2 !$OMP PARALLEL PRIVATE(i, j) i=3 j=j+2 !$OMP END PARALLEL PRINT *, i, j ! Output: 1 2 END PROGRAM
Chapter 7. Parallel programming with XL Fortran 163 Related reference: DO / END DO PARALLEL / END PARALLEL PARALLEL DO / END PARALLEL DO PARALLEL SECTIONS / END PARALLEL SECTIONS PARALLEL WORKSHARE / END PARALLEL WORKSHARE SECTIONS / END SECTIONS SINGLE / END SINGLE “TASK / END TASK” on page 133 REDUCTION Purpose
The REDUCTION clause updates named variables declared on the clause within the directive construct. Intermediate values of REDUCTION variables are not used within the parallel construct, other than in the updates themselves.
Syntax
REDUCTION ( variable_name_list ) op_fnc :
op_fnc is a reduction_op or a reduction_function that appears in all REDUCTION statements involving this variable. You must not specify more than one REDUCTION operator or function for a variable in the directive construct. To maintain OpenMP API compliance, you must specify op_fnc for the REDUCTION clause.
A REDUCTION statement can have one of the following forms:
reduction_var_ref = expr reduction_op reduction_var_ref
reduction_var_ref = reduction_var_ref reduction_op expr
reduction_var_ref = reduction_function (expr, reduction_var_ref)
reduction_var_ref = reduction_function (reduction_var_ref, expr)
where: reduction_var_ref is a variable or subobject of a variable that appears in a REDUCTION clause reduction_op is one of the intrinsic operators: +, -, *, .AND., .OR., .EQV., .NEQV.,or .XOR.
164 XL Fortran: Optimization and Programming Guide when reduction_op is an intrinsic operator, it should be the last operation performed on the right side. reduction_function is one of the intrinsic procedures: MAX, MIN, IAND, IOR,orIEOR. expr should not contain references to reduction_var_ref
The canonical initialization value of each of the operators and intrinsics are shown in the following table. The actual initialization value will be consistent with the data type of your corresponding REDUCTION variable.
Intrinsic Operator Initialization +0 *1 -0 .AND. .TRUE. .OR. .FALSE. .EQV. .TRUE. .NEQV. .FALSE. .XOR. .FALSE. Intrinsic Procedure Initialization MAX Smallest representable number MIN Largest representable number IAND All bits on IOR 0 IEOR 0
Rules
The following rules apply to REDUCTION statements: v A variable in the REDUCTION clause must only occur in a REDUCTION statement within the directive construct on which the REDUCTION clause appears. v The two reduction_var_refs that appear in a REDUCTION statement must be lexically identical. v You cannot use the following form of the REDUCTION statement: reduction_var_ref = expr operator reduction_var_ref, where operator is any operator other than reduction_op.
When you specify individual members of a common block in a REDUCTION clause, the storage of the specified variable is no longer associated with the storage of the common block.
Any variable you specify in a REDUCTION clause of a work-sharing construct must be shared in the enclosing PARALLEL construct.
A variable that appears in the REDUCTION clause of a parallel construct can also appear in a PRIVATE clause on a work-sharing construct.
Chapter 7. Parallel programming with XL Fortran 165 If you use a REDUCTION clause on a construct that has a NOWAIT clause, the REDUCTION variable remains undefined until a barrier synchronization has been performed to ensure that all threads have completed the REDUCTION clause.
A REDUCTION variable must not appear in a FIRSTPRIVATE, PRIVATE,or LASTPRIVATE clause of another construct within the dynamic extent of the construct in which it appeared as a REDUCTION variable.
If you specify op_fnc for the REDUCTION clause, each variable in the variable_name_list must be of intrinsic type. The variable can only appear in a REDUCTION statement within the lexical extent of the directive construct. You must specify op_fnc if the directive uses the trigger_constant $OMP.
The REDUCTION clause specifies named variables that appear in reduction operations. The compiler will maintain local copies of such variables, but will combine them upon exit from the construct. The intermediate values of the REDUCTION variables are combined in random order, dependent on which threads finish their calculations first. Therefore, there is no guarantee that bit-identical results will be obtained from one parallel run to another. This is true even if the parallel runs use the same number of threads, scheduling type, and chunk size.
Variables that you specify as REDUCTION or LASTPRIVATE to a parallel construct become defined at the end of the construct. If you have concurrent definitions or uses of REDUCTION or LASTPRIVATE variables on multiple threads, you must ensure that the threads are synchronized at the end of the construct when the variables become defined. For example, if multiple threads encounter a PARALLEL construct with a REDUCTION variable, you must synchronize the threads when they reach the END PARALLEL directive, because the REDUCTION variable becomes defined at END PARALLEL. Therefore the whole PARALLEL construct must be enclosed within a synchronization construct.
If an allocatable array appears on a REDUCTION clause, it must have an allocation status of allocated upon entrance into the construct that contains the REDUCTION clause. Additionally, the private copies of the REDUCTION variable must not be deallocated or allocated within the region.
A variable in the REDUCTION clause must be of intrinsic type. A variable in the REDUCTION clause, or any element thereof, must not be any of the following: v A pointee v An assumed-size array v A THREADLOCAL common block v A THREADPRIVATE common block or its members v A THREADPRIVATE variable v An allocatable scalar object v A Fortran 90 pointer
These rules describe the use of REDUCTION on OpenMP directives. If you are using the REDUCTION clause on the INDEPENDENT directive, see the INDEPENDENT directive in the XL Fortran Language Reference directive.
166 XL Fortran: Optimization and Programming Guide Example for thread-level speculative execution
The following example demonstrates the use of REDUCTION in SPECULATIVE directives. PROGRAM MAIN INTEGER(2), ALLOCATABLE :: i2a(:) INTEGER(2) :: i2b
ALLOCATE(i2a(5)) i2a = [5, 7, -5, 13, 9] i2b=6
!SEP$ SPECULATIVE DO REDUCTION(MAX: i2b), NUM_THREADS(4) DOi=1,5 IF(i2b == 6) STOP 11 i2b = MAX(i2b, i2a(i)) END DO !SEP$ END SPECULATIVE DO
IF(i2b /= 13) STOP 12 DEALLOCATE(i2a) END PROGRAM Related reference: DO / END DO PARALLEL / END PARALLEL PARALLEL DO / END PARALLEL DO PARALLEL SECTIONS / END PARALLEL SECTIONS PARALLEL WORKSHARE / END PARALLEL WORKSHARE SECTIONS / END SECTIONS SCHEDULE Purpose
You can use the SCHEDULE clause to specify the chunking method for parallelization. Work is assigned to threads in different manners depending on the scheduling type or chunk size used.
Syntax
SCHEDULE ( sched_type ) ,n
sched_type is one of AFFINITY, AUTO, DYNAMIC, GUIDED, RUNTIME,or STATIC.
Note: Thread-level speculative execution supports the STATIC scheduling type. You can also specify DYNAMIC, GUIDED, or other types, but the runtime treats all these values as static scheduling. n must be a positive scalar integer expression; do not specify n for the AUTO and RUNTIME schedule type. If you are using the trigger_constant $OMP, do not specify the scheduling type AFFINITY.
Chapter 7. Parallel programming with XL Fortran 167 AFFINITY The iterations of a loop are initially divided into number_of_threads partitions, containing CEILING(number_of_iterations / number_of_threads) iterations. Each partition is initially assigned to a thread, and is then further subdivided into chunks containing n iterations, if n has been specified. If n has not been specified, then the chunks consist of CEILING(number_of_iterations_remaining_in_partition / 2) loop iterations. When a thread becomes free, it takes the next chunk from its initially assigned partition. If there are no more chunks in that partition, then the thread takes the next available chunk from a partition that is initially assigned to another thread. Threads that are active will complete the work in a partition that is initially assigned to a sleeping thread. AUTO The compiler and runtime system choose the most appropriate mapping of iteration to threads for each loop. DYNAMIC If n has been specified, the iterations of a loop are divided into chunks containing n iterations each. If n has not been specified, then the default chunk size is 1 iteration. Threads are assigned these chunks on a "first-come, first-do" basis. Chunks of the remaining work are assigned to available threads, until all work has been assigned. If a thread is asleep, its assigned work will be taken over by an active thread, once that other thread becomes available. GUIDED If you specify a value for n, the iterations of a loop are divided into chunks such that the size of each successive chunk is exponentially decreasing. n specifies the size of the smallest chunk, except possibly the last. If you do not specify a value for n, the default value is 1. The size of the initial chunk is proportional to CEILING(number_of_iterations / number_of_threads) iterations. Subsequent chunks are proportional to CEILING(number_of_iterations_remaining / number_of_threads) iterations. If n is greater than 1, each chunk should contain fewer than n iterations (except for the last chunk to be assigned, which can have fewer than n iterations. As each thread finishes a chunk, it dynamically obtains the next available chunk. You can use guided scheduling in a situation in which multiple threads in a team might arrive at a DO work-sharing construct at varying times, and each iteration requires roughly the same amount of work. For example, if you have a DO loop preceded by one or more work-sharing SECTIONS or DO constructs with NOWAIT clauses, you can guarantee that no thread waits at the barrier longer than it takes another thread to execute its final iteration, or final k iterations if a chunk size of k is specified. The GUIDED schedule requires the fewest synchronizations of all the scheduling methods. An n expression is evaluated outside of the context of the DO construct. Any function reference in the n expression must not have side effects.
168 XL Fortran: Optimization and Programming Guide The value of the n parameter on the SCHEDULE clause must be the same for all of the threads in the team. RUNTIME Determine the scheduling type at run time. At run time, the scheduling type can be specified using the environment variable OMP_SCHEDULE. If no scheduling type is specified using that variable, the default scheduling type used is AUTO. STATIC If n has been specified, the iterations of a loop are divided into chunks that contain n iterations. Each thread is assigned chunks in a "round robin" fashion. This is known as block cyclic scheduling. If the value of n is 1, then the scheduling type is specifically referred to as cyclic scheduling. If n has not been specified, the chunks will contain CEILING(number_of_iterations / number_of_threads) iterations. Each thread is assigned one of these chunks. This is known as block cyclic scheduling. If a thread is asleep and it has been assigned work, it will be awakened so that it may complete its work. The STATIC schedule ensures that the same logical iteration numbers are assigned to threads in two work-sharing loop regions if the following conditions are satisfied: v Both loop regions have the same number of loop iterations v Both loop regions either have the same value of n specified, or have no n specified v Both loop regions bind to the same parallel region A data dependence between the same logical iterations in two such loops is guaranteed to be satisfied to allow the safe use of the NOWAIT clause. In addition, you must make sure that all three conditions mentioned above are satisfied to get the correct result. Consecutive loop constructs with STATIC schedule with NOWAIT clause now guarantee the same iterations are being assigned to the same thread in the constructs. For an example of the loop constructs that satisfy all three conditions, see “Example for OpenMP” on page 170.
Rules
You must not specify the SCHEDULE clause more than once for a particular DO directive.
Example for thread-level speculative execution
The following example demonstrates the use of SCHEDULE(STATIC, n) in SPECULATIVE directives. PROGRAM MAIN REAL(4), DIMENSION(200) :: x, y, z REAL(4), PARAMETER :: c1 = 1.2345_4, c2 = 5.6789_4 REAL(4), PARAMETER :: e = 2.718281828_4, pi = 3.141592654_4 INTEGER, PARAMETER ::n=6 INTEGER :: num, a, b
TYPE schedinfo
Chapter 7. Parallel programming with XL Fortran 169 CHARACTER(10) :: name INTEGER :: chunk(n) END TYPE TYPE(schedinfo) :: sc
sc = schedinfo(’static’, (/0, 10, 50, 100, 200, 0/)) sc%chunk(n) = MAX_INT
DOi=1,n num = sc%chunk(i) !SEP$ SPECULATIVE DO SCHEDULE(STATIC, num) DOj=1,200 x(j) = (c1 + j/10) * pi y(j) = (c2 + j/100) * e END DO !SEP$ END SPECULATIVE DO END DO END PROGRAM
Example for OpenMP
The following example illustrates loop constructs that satisfy all three conditions listed in the STATIC section. !$OMP PARALLEL !$OMP DO SCHEDULE(STATIC) DOi=1,n c(i) = (a(i) + b(i)) / 2.0; ENDDO !$OMP END DO NOWAIT
!$OMP DO SCHEDULE(STATIC) DOi=1,n z(i) = sqrt(c(i)) ENDDO !$OMP END DO !$OMP END PARALLEL Related reference: “DO / END DO” on page 100 “PARALLEL DO / END PARALLEL DO” on page 113 SHARED Purpose
All sections use the same copy of the variables and common blocks you specify in data_scope_entity_list.
The SHARED clause specifies variables that must be available to all threads. If you specify a variable as SHARED, you are stating that all threads can safely share a single copy of the variable.
Syntax
SHARED ( data_scope_entity_list )
data_scope_entity
170 XL Fortran: Optimization and Programming Guide named_variable / common_block_name /
named_variable is a named variable that is accessible in the directive construct common_block_name is a common block name that is accessible in the directive construct
Rules
A variable in the SHARED clause must not be either: v A pointee v A THREADLOCAL common block. v A THREADPRIVATE common block or its members. v A THREADPRIVATE variable.
If a SHARED variable, a subobject of a SHARED variable, or an object associated with a SHARED variable or subobject of a SHARED variable appears as an actual argument in a reference to a non-intrinsic procedure and: v The actual argument is an array section with a vector subscript; or v The actual argument is – An array section, – An assumed-shape array, or, – A pointer array and the associated dummy argument is an explicit-shape or assumed-size array; then any references to or definitions of the shared storage that is associated with the dummy argument by any other thread must be synchronized with the procedure reference. In other words, you must structure your code in such a way that if a thread encounters a procedure reference, then the procedure call by that thread and any reference to or definition of the shared storage by any other thread will always occur in the same sequence. You can do this, for example, by placing the procedure reference after a BARRIER.
Example for thread-level speculative execution
The following example demonstrates the use of SHARED in SPECULATIVE directives. PROGRAM MAIN REAL(8) :: rr8 INTEGER(8) :: i8
i8=1 rr8=0
!SEP$ SPECULATIVE DO SHARED(i8) SHARED(rr8) DOi=1,100 rr8 = rr8 + 1.55_8
Chapter 7. Parallel programming with XL Fortran 171 i8=i8*2_8 END DO !SEP$ END SPECULATIVE DO END PROGRAM
Example for OpenMP
In the following example, the procedure reference with an array section actual argument is required to be synchronized with references to the dummy argument by placing the procedure reference in a critical section, because the associated dummy argument is an explicit-shape array. INTEGER :: abc(10)
i=2 j=5
!$OMP PARALLEL DEFAULT(NONE), SHARED(abc, i, j) !$OMP CRITICAL ! Actual argument is an array section. ! The procedure reference must be in a critical section. CALL sub1(abc(i: j)) !$OMP END CRITICAL !$OMP END PARALLEL
CONTAINS SUBROUTINE sub1(arr) INTEGER:: arr(1: 4) DOi=1,4 arr(i) = i END DO END SUBROUTINE END Related reference: PARALLEL / END PARALLEL PARALLEL DO / END PARALLEL DO PARALLEL SECTIONS / END PARALLEL SECTIONS PARALLEL WORKSHARE / END PARALLEL WORKSHARE “TASK / END TASK” on page 133 UNTIED Purpose
The UNTIED clause is used with the TASK directive. When a task region is suspended, untied tasks can be resumed by any thread in a team.
Syntax
UNTIED
Rules
The UNTIED clause is ignored if either of the following conditions is true: v A FINAL clause is specified on the same task construct and the FINAL clause expression evaluates to .TRUE.. v The task is an included task.
172 XL Fortran: Optimization and Programming Guide Related reference: “TASK / END TASK” on page 133
Routines for parallel programming This section describes routines that are supported by OpenMP, POMP, thread-level speculative execution, and transactional memory on Blue Gene/Q platforms. Routines for OpenMP The OpenMP specification provides a number of routines that you can use to control and query the parallel execution environment, timing, and lock.
Parallel threads created by the runtime environment through the OpenMP interface are considered independent of the threads you create and control using calls to the Fortran Pthreads library module. References within the following descriptions to "serial portions of the program" refer to portions of the program that are executed by only one of the threads that have been created by the runtime environment. For example, you can create multiple threads by using f_pthread_create. However, if you then call omp_get_num_threads from outside of an OpenMP parallel block, or from within a serialized nested parallel region, the function will return 1, regardless of the number of threads that are currently executing.
OpenMP runtime library calls must not appear in PURE and ELEMENTAL procedures. Table 19. OpenMP execution environment routines
omp_get_active_level omp_get_thread_num omp_get_ancestor_thread_num omp_get_schedule omp_get_dynamic omp_get_team_size omp_get_level omp_get_thread_limit omp_get_max_active_levels omp_in_final omp_get_max_threads omp_in_parallel omp_get_nested omp_set_dynamic omp_get_num_procs omp_set_max_active_levels omp_get_num_threads omp_set_nested omp_set_num_threads omp_set_schedule
Included in the OpenMP runtime library are two routines that support a portable wall-clock timer. Table 20. OpenMP timing routines
omp_get_wtick omp_get_wtime
The OpenMP runtime library also supports a set of simple and nestable lock routines. You must only lock variables through these routines. Simple locks may not be locked if they are already in a locked state. Simple lock variables are associated with simple locks and may only be passed to simple lock routines. Nestable locks may be locked multiple times by the same thread. Nestable lock variables are associated with nestable locks and may only be passed to nestable lock routines. Note that locks are now associated with task regions, and no longer with threads as such, in accordance with changes in the OMP standard.
Chapter 7. Parallel programming with XL Fortran 173 For all the routines listed below, the lock variable is an integer whose KIND type parameter is denoted either by the symbolic constant omp_lock_kind,orby omp_nest_lock_kind.
This variable is set to '8' for 64-bit applications. Table 21. OpenMP simple lock routines
omp_destroy_lock omp_test_lock omp_init_lock omp_unset_lock omp_set_lock
Table 22. OpenMP nestable lock routines
omp_destroy_nest_lock omp_test_nest_lock omp_init_nest_lock omp_unset_nest_lock omp_set_nest_lock
Note: You can define and implement your own versions of the OpenMP routines. However, by default, the compiler will substitute the XL Fortran versions of the OpenMP routines regardless of the existence of other implementations, unless you specify the -qnoswapomp compiler option. For more information, see XL Fortran Compiler Reference. omp_destroy_lock(svar) Purpose
The omp_destroy_lock subroutine disassociates a given lock variable from all locks. You must use omp_init_lock to reinitialize a lock variable that was destroyed with a call to omp_destroy_lock before using it again as a lock variable.
If you call omp_destroy_lock with an uninitialized lock variable, the result of the call is undefined.
Class
Subroutine.
Argument Type and Attributes svar Type integer with kind omp_lock_kind.
Result Type and Attributes
None.
Result Value
None.
174 XL Fortran: Optimization and Programming Guide Examples
In the following example, threads and their associated tasks are generated by the parallel region, and one at a time, each task gains ownership of the lock associated with the lock variable LCK, prints the thread ID, and releases ownership of the lock. USE omp_lib INTEGER(kind=omp_lock_kind) LCK INTEGER ID CALL omp_init_lock(LCK) !$OMP PARALLEL SHARED(LCK), PRIVATE(ID) ID = omp_get_thread_num() CALL omp_set_lock(LCK) PRINT *,’MY THREAD ID IS’, ID CALL omp_unset_lock(LCK) !$OMP END PARALLEL CALL omp_destroy_lock(LCK) END omp_destroy_nest_lock(nvar) Purpose
The omp_destroy_nest_lock subroutine initializes a nestable lock variable, causing the lock variable to become undefined. The variable nvar must be an unlocked and initialized nestable lock variable.
If you call omp_destroy_nest_lock using an uninitialized variable, the result is undefined.
Class
Subroutine.
Argument Type and Attributes nvar Type integer with kind omp_nest_lock_kind.
Result Type and Attributes
None.
Result Value
None. omp_get_active_level() Purpose
The omp_get_active_level function returns the number of nested, active parallel regions.
Class
Function.
Argument Type and Attributes
None.
Chapter 7. Parallel programming with XL Fortran 175 Result Type and Attributes
Default integer.
Result Value
An integer that indicates the number of nested, active parallel regions. omp_get_ancestor_thread_num(level) Purpose
The omp_get_ancestor_thread_num function returns the thread number of the ancestor at a given nested level of the current thread.
Class
Function.
Argument Type and Attributes level Default integer.
Result Type and Attributes
Default integer.
Result Value
The thread number of the ancestor at a given nested level ( level ) of the current thread. If level is outside the range of 0 and the nested level of the current thread, as returned by the omp_get_level routine, the function returns -1. omp_get_dynamic() Purpose
The omp_get_dynamic function returns .TRUE. if dynamic thread adjustment by the runtime environment is enabled. Otherwise, the omp_get_dynamic function returns .FALSE.
Class
Function.
Argument Type and Attributes
None.
Result Type and Attributes
Default logical.
Result Value
.TRUE. if dynamic thread adjustment by the runtime environment is enabled; .FALSE. otherwise.
176 XL Fortran: Optimization and Programming Guide omp_get_level() Purpose
The omp_get_level function returns the number of nested parallel regions (both active and inactive).
Class
Function.
Argument Type and Attributes
None.
Result Type and Attributes
Default integer.
Result Value
The number of nested parallel regions (both active and inactive) in which the generating task is executing, not including the implicit parallel region. omp_get_max_active_levels() Purpose
The omp_get_max_active_levels function returns the maximum number of nested, active parallel regions.
Class
Function.
Argument Type and Attributes
None.
Result Type and Attributes
Default integer.
Result Value
The maximum number of nested, active parallel regions that is allowed.
Note: XL Fortran does not support OpenMP nested parallelism. This function always returns 1. omp_get_max_threads() Purpose
The omp_get_max_threads routine returns the first value of num_list for the OMP_NUM_THREADS environment variable. This value is the maximum number of threads that can be used to form a new team if a parallel region without a num_threads clause is encountered.
Chapter 7. Parallel programming with XL Fortran 177 If you use omp_set_num_threads to change the number of threads, subsequent calls to omp_get_max_threads will return the new value.
The routine has global scope, which means that the maximum value it returns applies to all routines, subroutines, and compilation units in the program. It returns the same value whether executing from a serial or parallel region.
You can use omp_get_max_threads to allocate maximum-sized data structures for each thread when you have enabled dynamic thread adjustment by passing omp_set_dynamic an argument which evaluates to .TRUE.
Class
Function.
Argument Type and Attributes
None.
Result Type and Attributes
Default integer.
Result Value
The maximum number of threads that can execute concurrently in a single parallel region. omp_get_nested() Purpose
The omp_get_nested function returns .TRUE. if nested parallelism is enabled and .FALSE. if nested parallelism is disabled.
Currently, XL Fortran does not support OpenMP nested parallelism.
Class
Function
Argument Type and Attributes
None.
Result Type and Attributes
Default logical.
Result Value
.TRUE. if nested parallelism is enabled. .FALSE. otherwise.
178 XL Fortran: Optimization and Programming Guide omp_get_num_procs() Purpose
The omp_get_num_procs function returns the number of online processors on the machine.
Class
Function.
Argument Type and Attributes
None.
Result Type and Attributes
Default integer.
Result Value
The number of online processors on the machine. omp_get_num_threads() Purpose
The omp_get_num_threads function returns the number of threads in the team currently executing the parallel region from which it is called. The function binds to the closest enclosing PARALLEL directive.
The omp_set_num_threads subroutine and the OMP_NUM_THREADS environment variable control the number of threads in a team. If you do not explicitly set the number of threads, the runtime environment will use the number of online processors on the machine by default. The number of online processors is less than or equal to the number of physical processors actually installed in a machine.
If you call omp_get_num_threads from a serial portion of your program or from a nested parallel region that is serialized, the function returns 1.
Class
Function.
Argument Type and Attributes
None.
Result Type and Attributes
Default integer.
Result Value
The number of threads in the team currently executing the parallel region from which the function is called.
Chapter 7. Parallel programming with XL Fortran 179 Examples USE omp_lib INTEGER N1, N2
N1 = omp_get_num_threads() PRINT *, N1 !$OMP PARALLEL PRIVATE(N2) N2 = omp_get_num_threads() PRINT *, N2 !$OMP END PARALLEL END
The omp_get_num_threads call returns 1 in the serial section of the code, so N1 is assigned the value 1. N2 is assigned the number of threads in the team executing the parallel region, so the output of the second print statement will be an arbitrary number less than or equal to the value returned by omp_get_max_threads. omp_get_schedule(kind, modifier) Purpose
The omp_get_schedule subroutine returns the scheduling type that is applied when using the runtime schedule. The argument kind returns the type of schedule that is used. modifier represents the chunk size that is set for applicable schedule types.
Class
Subroutine.
Argument Type and Attributes kind Integer of kind omp_sched_kind. The value returned for kind is one of the following constants that are defined in omp_lib module: v omp_sched_static v omp_sched_dynamic v omp_sched_guided v omp_sched_auto v omp_sched_affinity where omp_sched_affinity is not part of the OpenMP specification. modifier Default integer. For the schedule type dynamic, guided,orstatic, modifier is the chunk size that is set. For the schedule type auto, modifier has no meaning.
Result Type and Attributes
None.
Result Value
None.
180 XL Fortran: Optimization and Programming Guide omp_get_team_size(level) Purpose
The omp_get_team_size function returns the size of the thread team that the ancestor belongs to.
Class
Function.
Argument Type and Attributes level Default integer. level is the nested level of the current thread.
Result Type and Attributes
Default integer.
Result Value
The size of the thread team that the ancestor belongs to. If level is outside of the range of 0 and the nested level of the current thread, as returned by the omp_get_level function, the function returns -1. omp_get_thread_limit() Purpose
The omp_get_thread_limit function returns the maximum number of OpenMP threads that are available to the program.
Class
Function.
Argument Type and Attributes
None.
Result Type and Attributes
Default integer.
Result Value
The maximum number of OpenMP threads that are available to the program. omp_get_thread_num() Purpose
The omp_get_thread_num function returns the number of the currently executing thread within the team. The number returned will always be between 0 and NUM_PARTHDS -1.NUM_PARTHDS is the number of currently executing threads within the team. The master thread of the team returns a value of 0.
Chapter 7. Parallel programming with XL Fortran 181 If you call omp_get_thread_num from within a serial region, from within a serialized nested parallel region, or from outside the dynamic extent of any parallel region, this function will return a value of 0.
This function binds to the closest parallel region.
Class
Function.
Argument Type and Attributes
None.
Result Type and Attributes
Default integer.
Result Value
The value of the currently executing thread within the team between 0 and NUM_PARTHDS -1.NUM_PARTHDS is the number of currently executing threads within the team. A call to omp_get_thread_num from a serialized nested parallel region, or from outside the dynamic extent of any parallel region returns 0.
Examples
The following example illustrates the return value of the omp_get_thread_num routine in a PARALLEL region and a MASTER construct. USE omp_lib INTEGER NP call omp_set_num_threads(4) ! 4 threads are used in the ! parallel region
!$OMP PARALLEL PRIVATE(NP) NP = omp_get_thread_num() CALL WORK(’in parallel’, NP)
!$OMP MASTER NP = omp_get_thread_num() CALL WORK(’in master’, NP) !$OMP END MASTER !$OMP END PARALLEL END SUBROUTINE WORK(msg, THD_NUM) INTEGER THD_NUM character(*) msg PRINT *, msg, THD_NUM END
Output: in parallel 1 in parallel 3 in parallel 2 in parallel 0 in master 0
(The order may be different.)
182 XL Fortran: Optimization and Programming Guide omp_get_wtick() Purpose
The omp_get_wtick function returns a double precision value equal to the number of seconds between consecutive clock ticks.
Class
Function.
Argument Type and Attributes
None.
Result Type and Attributes
Double precision real.
Result Value
The number of seconds between consecutive ticks of the operating system real-time clock.
Examples USE omp_lib DOUBLE PRECISION WTICKS WTICKS = omp_get_wtick() PRINT *, ’The clock ticks ’, 10 / WTICKS, & ’ times in 10 seconds.’ END omp_get_wtime() Purpose
The omp_get_wtime function returns a double precision value equal to the number of seconds since the initial value of the operating system real-time clock. The initial value is guaranteed not to change during execution of the program.
The value returned by the omp_get_wtime function is not consistent across all threads in the team.
Class
Function.
Argument Type and Attributes
None.
Result Type and Attributes
Double precision real.
Result Value
The number of seconds since the initial value of the operating system real-time clock.
Chapter 7. Parallel programming with XL Fortran 183 Examples USE omp_lib DOUBLE PRECISION START, END START = omp_get_wtime() ! Work to be timed END = omp_get_wtime() PRINT *, ’Stuff took ’, END - START, ’ seconds.’ END omp_in_final() Purpose
The omp_in_final routine returns .TRUE. if the routine is called in a final task region. Otherwise, the routine returns .FALSE..
Class
Function.
Argument Type and Attributes
None.
Result Type and Attributes
Default logical.
Result Value
If the routine is called in a final task region, the result value is .TRUE.; otherwise, the result value is .FALSE.. omp_in_parallel() Purpose
The omp_in_parallel function returns .TRUE. if you call it from the dynamic extent of a region executing in parallel and returns .FALSE. otherwise. If you call omp_in_parallel from a region that is serialized but nested within the dynamic extent of a region executing in parallel, the function will still return .TRUE.. (Nested parallel regions are serialized by default. See “omp_set_nested(enable_expr)” on page 189 and the OMP_NESTED environment variable for more information.)
Class
Function.
Argument Type and Attributes
None.
Result Type and Attributes
Default logical.
184 XL Fortran: Optimization and Programming Guide Result Value
.TRUE. if called from the dynamic extent of a region executing in parallel. .FALSE. otherwise.
Examples
In the following example, the first call to omp_in_parallel returns .FALSE. because the call is outside the dynamic extent of any parallel region. The second call returns .TRUE., even if the nested PARALLEL DO loop is serialized, because the call is still inside the dynamic extent of the outer PARALLEL DO loop. USE omp_lib INTEGER N, M N=4 M=3 PRINT*, omp_in_parallel() !$OMP PARALLEL DO DO I = 1,N !$OMP PARALLEL DO DO J=1, M PRINT *, omp_in_parallel() END DO !$OMP END PARALLEL DO END DO !$OMP END PARALLEL DO END omp_init_lock(svar) Purpose
The omp_init_lock subroutine initializes a lock and associates it with the lock variable passed in as a parameter. After the call to omp_init_lock, the initial state of the lock variable is unlocked.
If you call this routine with a lock variable that you have already initialized, the result of the call is undefined.
Class
Subroutine.
Argument Type and Attributes svar Integer of kind omp_lock_kind.
Result Type and Attributes
None.
Result Value
None.
Examples
In the following example, threads and their associated tasks are generated by the parallel region, and one at a time, each task gains ownership of the lock associated with the lock variable LCK, prints the thread ID, and releases ownership of the lock.
Chapter 7. Parallel programming with XL Fortran 185 USE omp_lib INTEGER(kind=omp_lock_kind) LCK INTEGER ID CALL omp_init_lock(LCK) !$OMP PARALLEL SHARED(LCK), PRIVATE(ID) ID = omp_get_thread_num() CALL omp_set_lock(LCK) PRINT *,’MY THREAD ID IS’, ID CALL omp_unset_lock(LCK) !$OMP END PARALLEL CALL omp_destroy_lock(LCK) END omp_init_nest_lock(nvar) Purpose
The omp_init_nest_lock subroutine allows you to initialize a nestable lock and associate it with the lock variable you specify. The initial state of the lock variable is unlocked, and the initial nesting count is zero. The value of nvar must be an unitialized nestable lock variable.
If you call omp_init_nest_lock using a variable that is already initialized, the result is undefined.
Class
Subroutine.
Argument Type and Attributes nvar Integer of kind omp_nest_lock_kind.
Result Type and Attributes
None.
Result Value
None.
Examples
The following example illustrates the use of a nestable lock for updating variable P in the PARALLEL SECTIONS construct. USE omp_lib INTEGER P INTEGER A INTEGER B INTEGER ( kind=omp_nest_lock_kind ) LCK CALL omp_init_nest_lock ( LCK ) ! initialize the nestable lock !$OMP PARALLEL SECTIONS !$OMP SECTION CALL omp_set_nest_lock ( LCK ) P=P+A CALL omp_set_nest_lock ( LCK ) P=P+B CALL omp_unset_nest_lock ( LCK ) CALL omp_unset_nest_lock ( LCK ) !$OMP SECTION CALL omp_set_nest_lock ( LCK ) P=P+B
186 XL Fortran: Optimization and Programming Guide CALL omp_unset_nest_lock ( LCK ) !$OMP END PARALLEL SECTIONS
CALL omp_destroy_nest_lock ( LCK ) END omp_set_dynamic(enable_expr) Purpose
The omp_set_dynamic subroutine enables or disables dynamic adjustment, by the runtime environment, of the number of threads available to execute parallel regions.
If enable_expr is evaluated to .TRUE., the runtime environment can automatically adjust the number of threads that are used to execute subsequent parallel regions to obtain the best use of system resources. The number of threads you specify using omp_set_num_threads becomes the maximum, not exact, thread count.
If enable_expr is evaluated to .FALSE., dynamic adjustment of the number of threads is disabled. The runtime environment cannot automatically adjust the number of threads used to execute subsequent parallel regions. The value you pass to omp_set_num_threads becomes the exact thread count.
By default, dynamic thread adjustment is disabled. If your code depends on a specific number of threads for correct execution, you should explicitly disable dynamic threads.
If the routine is called from a portion of the program where the omp_in_parallel routine returns .TRUE., the routine has no effect.
This subroutine has precedence over the OMP_DYNAMIC environment variable.
Class
Subroutine.
Argument Type and Attributes enable_expr Logical.
Result Type and Attributes
None.
Result Value
None. omp_set_lock(svar) Purpose
The omp_set_lock subroutine forces the calling task region to wait until the specified lock is available before executing subsequent instructions. The calling task region is given ownership of the lock when it becomes available.
Chapter 7. Parallel programming with XL Fortran 187 If you call this routine with an uninitialized lock variable, the result of the call is undefined. If a task region that owns a lock tries to lock it again by issuing a call to omp_set_lock, the call produces a deadlock.
Class
Subroutine.
Argument Type and Attributes svar Integer of kind omp_lock_kind.
Result Type and Attributes
None.
Result Value
None.
Examples
In the following example, the lock variable LCK_X is used to avoid race conditions when updating the shared variable X. By setting the lock before each update to X and unsetting it after the update, you ensure that only one task region updates X at a given time. USE omp_lib INTEGER A(100), X INTEGER(kind=omp_lock_kind) LCK_X X=1 CALL omp_init_lock (LCK_X) !$OMP PARALLEL PRIVATE (I), SHARED (A, X) !$OMP DO DOI=3,100 A(I)=I*10 CALL omp_set_lock (LCK_X) X=X+A(I) CALL omp_unset_lock (LCK_X) END DO !$OMP END DO !$OMP END PARALLEL CALL omp_destroy_lock (LCK_X) END omp_set_max_active_levels(max_levels) Purpose
The omp_set_max_active_levels subroutine limits the number of nested, active parallel regions.
Class
Subroutine.
Argument Type and Attributes max_levels Default integer.
188 XL Fortran: Optimization and Programming Guide Result Type and Attributes
None.
Result Value
None. omp_set_nested(enable_expr) Purpose
The omp_set_nested subroutine enables or disables nested parallelism.
If enable_expr is evaluated to .FALSE., nested parallelism is disabled. Nested parallel regions are serialized, and they are executed by the current thread. This is the default setting.
If enable_expr is evaluated to .TRUE., nested parallelism is enabled. Parallel regions that are nested can deploy additional threads to the team. It is up to the runtime environment to determine whether additional threads should be deployed. Therefore, the number of threads used to execute parallel regions may vary from one nested region to the next.
If the routine is called from a portion of the program where the omp_in_parallel routine returns true, the routine has no effect.
This subroutine takes precedence over the OMP_NESTED environment variable.
Currently, XL Fortran does not support OpenMP nested parallelism.
Class
Subroutine.
Argument Type and Attributes enable_expr Logical.
Result Type and Attributes
Default logical.
Result Value
None. omp_set_nest_lock(nvar) Purpose
The omp_set_nest_lock subroutine allows you to set a nestable lock. The task region executing the subroutine will wait until the lock becomes available and then set that lock, incrementing the nesting count. A nestable lock is available if it is owned by the task region executing the subroutine, or is unlocked.
Chapter 7. Parallel programming with XL Fortran 189 Class
Subroutine.
Argument Type and Attributes nvar Integer of kind omp_nest_lock_kind.
Result Type and Attributes
None.
Result Value
None.
Examples USE omp_lib INTEGER P INTEGER A INTEGER B INTEGER ( kind=omp_nest_lock_kind ) LCK
CALL omp_init_nest_lock ( LCK )
!$OMP PARALLEL SECTIONS !$OMP SECTION CALL omp_set_nest_lock ( LCK ) P=P+A CALL omp_set_nest_lock ( LCK ) P=P+B CALL omp_unset_nest_lock ( LCK ) CALL omp_unset_nest_lock ( LCK ) !$OMP SECTION CALL omp_set_nest_lock ( LCK ) P=P+B CALL omp_unset_nest_lock ( LCK ) !$OMP END PARALLEL SECTIONS
CALL omp_destroy_nest_lock ( LCK ) END omp_set_num_threads(number_of_threads_expr) Purpose
The omp_set_num_threads routine specifies the number of threads to use for the next parallel region by setting the first value of num_list for the OMP_NUM_THREADS environment variable.
The number_of_threads_expr argument is evaluated, and its value is used as the number of threads. If you have enabled dynamic adjustment of the number of threads (see “omp_set_dynamic(enable_expr)” on page 187), omp_set_num_threads sets the maximum number of threads to use for the next parallel region. The runtime environment then determines the exact number of threads to use. However, when dynamic adjustment of the number of threads is disabled, omp_set_num_threads sets the exact number of threads to use in the next parallel region. If the number of threads you request exceeds the number your execution environment can support, your application will terminate.
190 XL Fortran: Optimization and Programming Guide This subroutine takes precedence over the OMP_NUM_THREADS environment variable.
If you call this subroutine from the dynamic extent of a region executing in parallel, the behavior of the subroutine is undefined.
Class
Subroutine.
Argument Type and Attributes number_of_threads_expr integer
Result Type and Attributes
None.
Result Value
None. omp_set_schedule(kind, modifier) Purpose
The omp_set_schedule routine affects the schedule that is applied when runtime is used as schedule kind. Use omp_set_schedule if you want to set the schedule type separately from the OMP_SCHEDULE environment variable.
Note: You can use the omp_get_schedule to return scheduling type. For details, see omp_get_schedule .
Class
Subroutine.
Argument Type and Attributes kind Type integer with kind omp_sched_kind. Must be one of the schedule types as represented by the following constants: v omp_sched_static v omp_sched_dynamic v omp_sched_guided v omp_sched_auto v omp_sched_affinity where omp_sched_affinity is not part of the OpenMP specification. modifier Default integer. For the schedule type dynamic, guided,orstatic, modifier is the chunk size that you want to set. Typically, it is a positive integer. If the value is less than one, the default is used. For the schedule type auto, modifier has no meaning. For the default setting of each schedule type, see -qsmp in the XL Fortran Compiler Reference.
Chapter 7. Parallel programming with XL Fortran 191 Result Type and Attributes
None.
Result Value
None. omp_test_lock(svar) Purpose
The omp_test_lock function attempts to set the lock associated with the specified lock variable. It returns .TRUE. if it was able to set the lock and .FALSE. otherwise. In either case, the calling task region will continue to execute subsequent instructions in the program.
If you call omp_test_lock with an uninitialized lock variable, the result of the call is undefined.
Class
Function.
Argument Type and Attributes svar Integer of kind omp_lock_kind.
Result Type and Attributes
Default logical.
Result Value
.TRUE. if the function was able to set the lock. .FALSE. otherwise.
Examples
In the following example, a task region repeatedly executes WORK_A until it can set the lock variable, LCK. When the lock is set, the task region executes WORK_B. USE omp_lib INTEGER LCK INTEGER ID CALL omp_init_lock (LCK) !$OMP PARALLEL SHARED(LCK), PRIVATE(ID) ID = omp_get_thread_num() DO WHILE (.NOT. omp_test_lock(LCK)) CALL WORK_A (ID) END DO CALL WORK_B (ID) CALL omp_unset_lock (LCK) !$OMP END PARALLEL CALL omp_destroy_lock (LCK) END omp_test_nest_lock(nvar) Purpose
The omp_test_nest_lock subroutine allows you to attempt to set a lock using the same method as omp_set_nest_lock, but the execution task region does not wait
192 XL Fortran: Optimization and Programming Guide for confirmation that the lock is available. If the lock is successfully set, the function will increment the nesting count and return the new nesting count. If the lock is unavailable the function returns a value of zero. Also, a child task sees a value of zero if the parent task has already set the same lock. The result value is always a default integer.
Class
Function.
Argument Type and Attributes nvar Integer of kind omp_nest_lock_kind.
Result Type and Attributes
Default integer.
Result Value
The new nesting count if the lock is successfully set; otherwise, it returns zero. omp_unset_lock(svar) Purpose
The omp_unset_lock subroutine causes the executing task region to release ownership of the specified lock. The lock can then be set by another task region as required. The behavior of the omp_unset_lock subroutine is undefined if either of the following conditions occur: v The calling task region does not own the lock specified. v The routine is called with an uninitialized lock variable.
Class
Subroutine.
Argument Type and Attributes svar Integer of kind omp_lock_kind.
Result Type and Attributes
None.
Result Value
None.
Examples USE omp_lib INTEGER A(100) INTEGER(kind=omp_lock_kind) LCK_X CALL omp_init_lock (LCK_X) !$OMP PARALLEL PRIVATE (I), SHARED (A, X) !$OMP DO DOI=3,100 A(I)=I*10 CALL omp_set_lock (LCK_X)
Chapter 7. Parallel programming with XL Fortran 193 X=X+A(I) CALL omp_unset_lock (LCK_X) END DO !$OMP END DO !$OMP END PARALLEL CALL omp_destroy_lock (LCK_X) END
In this example, the lock variable LCK_X is used to avoid race conditions when updating the shared variable X. By setting the lock before each update to X and unsetting it after the update, you ensure that only one task region is updating X at a given time. omp_unset_nest_lock(nvar) Purpose
The omp_unset_nest_lock subroutine allows you to release ownership of a nestable lock. The subroutine decrements the nesting count and releases the associated task region from ownership of the nestable lock.
Class
Subroutine.
Argument Type and Attributes nvar Integer of kind omp_lock_kind.
Result Type and Attributes
None.
Result Value
None.
Examples USE omp_lib INTEGER P INTEGER A INTEGER B INTEGER ( kind=omp_nest_lock_kind ) LCK
CALL omp_init_nest_lock ( LCK )
!$OMP PARALLEL SECTIONS !$OMP SECTION CALL omp_set_nest_lock ( LCK ) P=P+A CALL omp_set_nest_lock ( LCK ) P=P+B CALL omp_unset_nest_lock ( LCK ) CALL omp_unset_nest_lock ( LCK ) !$OMP SECTION CALL omp_set_nest_lock ( LCK ) P=P+B CALL omp_unset_nest_lock ( LCK ) !$OMP END PARALLEL SECTIONS
CALL omp_destroy_nest_lock ( LCK ) END
194 XL Fortran: Optimization and Programming Guide Routines for thread-level speculative execution To call the routines that are specific to thread-level speculative execution, you must use the speculation_util intrinsic module.
The routines for thread-level speculative execution have no effect unless the thread-level speculative execution is enabled with the "-qsmp=speculative" compiler option. Structure of statistic counters
The following structure that contains the statistic counters can be used with the thread-level speculative execution routines. It is declared in the speculation_util intrinsic module. TYPE SeReport
SEQUENCE
! Total number of chunks that are committed by nonespeculative threads INTEGER(8) totalNONSpecCommitted
! Total number of chunks that are committed by speculative threads INTEGER(8) totalSpecCommitted
! Total number of rollbacks for speculative threads INTEGER(8) totalRollbacks
! Total number of serialization caused by JMV conflicts INTEGER(8) totalSerializedJMV
! Total number of serialization caused by SE_MAX_NUM_ROLLBACK reached INTEGER(8) totalSerializedMAXRB
! Total number of serialization caused by incorrect hardware ! speculative mode or nesting INTEGER(8) totalSerializedOTHER
END TYPE se_get_all_stats(stats) Purpose
With the se_get_all_stats routine, you can retrieve statistical information for the thread-level speculative execution of all threads in a program.
Class
Subroutine.
Argument Type and Attributes stats An INTENT(OUT) SeReport variable.
Result Type and Attributes
None.
Result Value
None.
Chapter 7. Parallel programming with XL Fortran 195 Usage
When the se_get_all_stats routine is called, it updates the argument with cumulative statistics of all the regions of thread-level speculative execution that all hardware threads have run.
To use the se_get_all_stats routine, you must set the SE_REPORT_STAT_ENABLE environment variable to YES.
You must use this routine outside of parallel regions.
Related information v Environment variables for thread-level speculative execution se_print_stats()
Purpose
With the se_print_stats routine, you can write statistics about thread-level speculative execution to a log file.
Class
Subroutine.
Example ... !SEP$ SPECULATIVE DO DO-LOOP
CALL SE_PRINT_STATS() ...
Usage
After a region of thread-level speculative execution, you can call the se_print_stats routine to write cumulative statistics of all the regions of thread-level speculative execution that each hardware thread has run to a log file.
To enable the statistics log file to be generated at each call to the se_print_stats routine, you must set the SE_REPORT_LOG environment variable to FUNC, ALL, or VERBOSE.
By default, the log file is named se_report.log.rank where rank is the MPI rank of the process that called the se_print_stats routine. The file is saved in the current working directory of the program that uses thread-level speculative execution. To override the defaults, you can specify the SE_REPORT_NAME environment variable.
You must use this routine outside of parallel regions.
Note: If you use this function in a region of thread-level speculative execution, the code of that region is run in irrevocable mode.
Related information v Environment variables for thread-level speculative execution
196 XL Fortran: Optimization and Programming Guide reset_speculation_mode()
Purpose
With the reset_speculation_mode routine, you can switch mode in between thread-level speculative execution (SE) and transactional memory (TM) at runtime.
Class
Subroutine.
Example ... !SEP$ SPECULATIVE DO do-loop !SEP$ END SPECULATIVE DO
CALL RESET_SPECULATION_MODE()
!TM$ TM_ATOMIC code-block !TM$ END TM_ATOMIC ...
Usage
The call to the reset_speculation_mode routine is needed only when you are switching mode between TM and SE.
If your program consists of both TM and SE regions, it is recommended that you make a call to the reset_speculation_mode routine; otherwise, undefined behavior might be caused.
When you call the reset_speculation_mode routine, make sure all the TM or SE executions before the call are completed.
Related information v “Nesting OpenMP, transactional memory, and thread-level speculative execution” on page 90 Routines for transactional memory To call the routines that are specific to transactional memory, you must use the speculation_util intrinsic module.
The routines for transactional memory have no effect unless transactional memory is enabled with the "-qtm" compiler option. Structure of statistic counters
The following structure that contains the statistic counters can be used with the transactional memory routines. It is declared in the speculation_util intrinsic module. TYPE TmReport
SEQUENCE
! Thread ID
Chapter 7. Parallel programming with XL Fortran 197 INTEGER(8) hwThreadId
! Total number of transactions INTEGER(8) totalTransactions
! Total number of rollbacks for transactional memory threads INTEGER(8) totalRollbacks
! Total number of serialization caused by JMV conflicts INTEGER(8) totalSerializedJMV
! Total number of serialization caused by TM_MAX_NUM_ROLLBACK reached INTEGER(8) totalSerializedMAXRB
! Total number of serialization caused by incorrect hardware ! speculative mode or nesting INTEGER(8) totalSerializedOTHER
END TYPE tm_get_stats(stats) Purpose
With the tm_get_stats routine, you can retrieve statistical information for the transactional memory of a particular hardware thread in your program.
Class
Subroutine.
Argument Type and Attributes stats An INTENT(OUT) TmReport variable.
Result Type and Attributes
None.
Result Value
None.
Usage
When the tm_get_stats routine is called, it updates the argument with cumulative statistics of all the transactional atomic regions that a particular hardware thread has run.
To use the tm_get_stats routine, you must set the TM_REPORT_STAT_ENABLE environment variable to YES.
Related information v Environment variables for transactional memory
198 XL Fortran: Optimization and Programming Guide tm_get_all_stats(stats) Purpose
With the tm_get_all_stats routine, you can retrieve statistical information for the transactional memory of all hardware threads in a program.
Class
Subroutine.
Argument Type and Attributes stats An INTENT(OUT) TmReport variable.
Result Type and Attributes
None.
Result Value
None.
Usage
When the tm_get_all_stats routine is called, it updates the argument with cumulative statistics of all the transactional atomic regions that all hardware threads have run.
To use the tm_get_all_stats routine, you must set the TM_REPORT_STAT_ENABLE environment variable to YES.
You must use this routine outside of parallel regions.
Related information v Environment variables for transactional memory tm_print_stats() Purpose
With the tm_print_stats routine, you can write statistics for the transactional memory of a particular hardware thread to a log file.
Class
Subroutine.
Example ... !TM$ TM_ATOMIC code-block !TM$ END TM_ATOMIC
CALL TM_PRINT_STATS() ...
Chapter 7. Parallel programming with XL Fortran 199 Usage
After a transactional atomic region, you can call the tm_print_stats routine to write cumulative statistics of all the transactional atomic regions that a particular hardware thread has run to a log file.
To enable the statistics log file to be generated at each call to the tm_print_stats routine, you must set the TM_REPORT_LOG variable to FUNC, ALL, or VERBOSE.
By default, the log file is named tm_report.log.rank, where rank in the extension is the MPI rank of the process that called tm_print_stats. The log file is saved in the current working directory of the program that uses transactional memory. To override the defaults, you can specify the TM_REPORT_NAME environment variable.
Note: If you use this routine in a transactional atomic region, it causes the transaction running in irrevocable mode.
Related information v Environment variables for transactional memory tm_print_all_stats() Purpose
With the tm_print_all_stats routine, you can write statistical information for the transactional memory of all hardware threads in a program to a log file.
Class
Subroutine.
Example ... !TM$ TM_ATOMIC code-block !TM$ END TM_ATOMIC
CALL TM_PRINT_ALL_STATS() ...
Usage
When the tm_print_all_stats routine is called, it writes the cumulative statistics of all the transactional atomic regions that all hardware threads have run to a log file.
By default, the log file is named tm_report.log.rank, where rank in the extension is the MPI rank of the process that called tm_print_all_stats. The log file is saved in the current working directory of the program that uses transactional memory. To override the defaults, you can specify the TM_REPORT_NAME environment variable.
To enable the statistics log file to be generated at each call to the tm_print_all_stats routine, you must set the TM_REPORT_LOG variable to FUNC, ALL, or VERBOSE.
You must use this routine outside of parallel regions.
200 XL Fortran: Optimization and Programming Guide Related information v Environment variables for transactional memory reset_speculation_mode()
Purpose
With the reset_speculation_mode routine, you can switch mode in between thread-level speculative execution (SE) and transactional memory (TM) at runtime.
Class
Subroutine.
Example ... !SEP$ SPECULATIVE DO do-loop !SEP$ END SPECULATIVE DO
CALL RESET_SPECULATION_MODE()
!TM$ TM_ATOMIC code-block !TM$ END TM_ATOMIC ...
Usage
The call to the reset_speculation_mode routine is needed only when you are switching mode between TM and SE.
If your program consists of both TM and SE regions, it is recommended that you make a call to the reset_speculation_mode routine; otherwise, undefined behavior might be caused.
When you call the reset_speculation_mode routine, make sure all the TM or SE executions before the call are completed.
Related information v “Nesting OpenMP, transactional memory, and thread-level speculative execution” on page 90
Pthreads library module The Pthreads Library Module (f_pthread) is a Fortran 90 module that defines data types and routines to make it easier to interface with the Blue Gene/Q pthreads library. The Blue Gene/Q pthreads library is used to parallelize and to make your code thread-safe.
The f_pthread library module naming convention is the use of the prefix f_ before the corresponding Blue Gene/Q pthreads library routine name or type definition name.
In general, there is a one-to-one corresponding relationship between the procedures in the Fortran 90 module f_pthread and the library routines contained in the Blue Gene/Q pthreads library. However, some of the pthread routines have no
Chapter 7. Parallel programming with XL Fortran 201 corresponding procedures in this module because they are not supported on Blue Gene/Q. One example of these routines is the thread stack address option. There are also some non-pthread interfacing routines contained in the f_pthread library module. The f_maketime routine is one example and is included to return an absolute time in a f_timespec derived type variable.
Most of the routines return an integer value. A return value of 0 will always indicate that the routine call did not result in any error. Any non-zero return value indicates an error. Each error code has a corresponding definition of a system error code in Fortran. These error codes are available as Fortran integer constants. The naming of these error codes in Fortran is consistent with the corresponding Blue Gene/Q error code names. For example, EINVAL is the Fortran constant name of the error code EINVAL on the system. For a complete list of these error codes, refer to the file /usr/include/errno.h.
Note: The pthread module in XL Fortran is an extension to the standard Fortran language. Pthreads data structures, functions, and subroutines Pthreads Data Types v f_pthread_attr_t v f_pthread_cond_t v f_pthread_condattr_t v f_pthread_key_t v f_pthread_mutex_t v f_pthread_mutexattr_t v f_pthread_once_t v f_pthread_rwlock_t v f_pthread_rwlockattr_t v f_pthread_t v f_sched_param v f_timespec Functions that perform operations on thread attribute objects v f_pthread_attr_destroy(attr) v f_pthread_attr_getdetachstate(attr, detach) v f_pthread_attr_getguardsize(attr, guardsize) v f_pthread_attr_getinheritsched(attr, inherit) v f_pthread_attr_getschedparam(attr, param) v f_pthread_attr_getschedpolicy(attr, policy) v f_pthread_attr_getscope(attr, scope) v f_pthread_attr_getstack(attr, stackaddr, ssize) v f_pthread_attr_init(attr) v f_pthread_attr_setdetachstate(attr, detach) v f_pthread_attr_setguardsize(attr, guardsize) v f_pthread_attr_setinheritsched(attr, inherit) v f_pthread_attr_setschedparam(attr, param) v f_pthread_attr_setschedpolicy(attr, policy) v f_pthread_attr_setscope(attr, scope)
202 XL Fortran: Optimization and Programming Guide v f_pthread_attr_setstack(attr, stackaddr, ssize) Functions and Subroutines That Perform Operations on Threads v f_pthread_cancel(thread) v f_pthread_cleanup_pop(exec) v f_pthread_cleanup_push(cleanup, flag, arg) v f_pthread_create(thread, attr, flag, ent, arg) v f_pthread_detach(thread) v f_pthread_equal(thread1, thread2) v f_pthread_exit(ret) v f_pthread_getconcurrency() v f_pthread_getschedparam(thread, policy, param) v f_pthread_join(thread, ret) v f_pthread_kill(thread, sig) v f_pthread_self() v f_pthread_setconcurrency(new_level) v f_pthread_setschedparam(thread, policy, param) Functions that perform operations on mutex attribute objects v f_pthread_mutexattr_destroy(mattr) v f_pthread_mutexattr_getpshared(mattr, pshared) v f_pthread_mutexattr_gettype(mattr, type) v f_pthread_mutexattr_init(mattr) v f_pthread_mutexattr_setpshared(mattr, pshared) v f_pthread_mutexattr_settype(mattr, type) Functions that perform operations on mutex objects v f_pthread_mutex_destroy(mutex) v f_pthread_mutex_init(mutex, mattr) v f_pthread_mutex_lock(mutex) v f_pthread_mutex_trylock(mutex) v f_pthread_mutex_unlock(mutex) Functions that perform operations on attribute objects of condition variables v f_pthread_condattr_destroy(cattr) v f_pthread_condattr_getpshared(cattr, pshared) v f_pthread_condattr_init(cattr) v f_pthread_condattr_setpshared(cattr, pshared) Functions that perform operations on condition variable objects v f_maketime(delay) v f_pthread_cond_broadcast(cond) v f_pthread_cond_destroy(cond) v f_pthread_cond_init(cond, cattr) v f_pthread_cond_signal(cond) v f_pthread_cond_timedwait(cond, mutex, timeout)
Chapter 7. Parallel programming with XL Fortran 203 v f_pthread_cond_wait(cond, mutex) Functions that perform operations on thread-specific data v f_pthread_getspecific(key, arg) v f_pthread_key_create(key, dtr) v f_pthread_key_delete(key) v f_pthread_setspecific(key, arg) Functions and subroutines that perform operations to control thread cancelability v f_pthread_setcancelstate(state, oldstate) v f_pthread_setcanceltype(type, oldtype) v f_pthread_testcancel() Functions that perform operations on read-write lock attribute objects v f_pthread_rwlockattr_destroy(rwattr) v f_pthread_rwlockattr_getpshared(rwattr, pshared) v f_pthread_rwlockattr_init(rwattr) v f_pthread_rwlockattr_setpshared(rwattr, pshared) Functions that perform operations on read-write lock objects v f_pthread_rwlock_destroy(rwlock) v f_pthread_rwlock_init(rwlock, rwattr) v f_pthread_rwlock_rdlock(rwlock) v f_pthread_rwlock_tryrdlock(rwlock) v f_pthread_rwlock_trywrlock(rwlock) v f_pthread_rwlock_unlock(rwlock) v f_pthread_rwlock_wrlock(rwlock) Functions that perform operations for one-time initialization v f_pthread_once(once, initr) f_maketime(delay) Purpose
This function accepts an integer value specifying a delay in seconds and returns an f_timespec type object containing the absolute time, which is delay seconds from the calling moment. Class
Function Argument Type and Attributes delay INTEGER(4), INTENT(IN) Result Type and Attributes
TYPE (f_timespec)
204 XL Fortran: Optimization and Programming Guide Result Value
The absolute time, which is delay seconds from the calling moment, is returned. f_pthread_attr_destroy(attr) Purpose
This function must be called to destroy any previously initialized thread attribute objects when they will no longer be used. Threads that were created with this attribute object will not be affected in any way by this action. Memory that was allocated when it was initialized will be recollected by the system. Class
Function Argument Type and Attributes attr TYPE(f_pthread_attr_t), INTENT(IN) Result Type and Attributes
INTEGER(4) Result Value
On successful completion, this function returns 0. Otherwise, this function returns the following error. EINVAL The argument attr is invalid. f_pthread_attr_getdetachstate(attr, detach) Purpose
This function can be used to query the setting of the detach state attribute in the thread attribute object attr. The current setting will be returned through argument detach. Class
Function Argument Type and Attributes attr TYPE(f_pthread_attr_t), INTENT(IN) detach INTEGER(4), INTENT(OUT) Contains one of the following values: PTHREAD_CREATE_DETACHED: when a thread attribute object of this attribute setting is used to create a new thread, the newly created thread will be in detached state.
Chapter 7. Parallel programming with XL Fortran 205 PTHREAD_CREATE_JOINABLE: when a thread attribute object of this attribute setting is used to create a new thread, the newly created thread will be in undetached state. Result Type and Attributes
INTEGER(4) Result Value
On successful completion, this function returns 0. Otherwise, this function returns the following error: EINVAL The argument attr is invalid. f_pthread_attr_getguardsize(attr, guardsize) Purpose
This function is used to get the guardsize attribute in the thread attribute object attr. The current setting of the attribute will be returned through the argument guardsize. Class
Function Argument Type and Attributes attr TYPE(f_pthread_attr_t), INTENT(IN) guardsize INTEGER(KIND=register_size), INTENT(IN) where register_size is 8 in 64-bit mode. Result Type and Attributes
INTEGER(4) Result Value
On successful completion, this function returns 0. Otherwise, this function returns the following error: EINVAL The argument attr is invalid. f_pthread_attr_getinheritsched(attr, inherit) Purpose
This function can be used to query the inheritance scheduling attribute in the thread attribute object attr. The current setting will be returned through the argument inherit.
206 XL Fortran: Optimization and Programming Guide Class
Function Argument Type and Attributes attr TYPE(f_pthread_attr_t), INTENT(OUT) inherit INTEGER(4) On return from the function, inherit contains one of the following values: PTHREAD_INHERIT_SCHED: indicating that newly created threads will inherit the scheduling property of the parent thread and ignore the scheduling property of the thread attribute object used to create them. PTHREAD_EXPLICIT_SCHED: the scheduling property in the thread attribute object will be assigned to the newly created threads when it is used to create them. Result Type and Attributes
INTEGER(4) Result Value
On successful completion, this function returns 0. Otherwise this function returns the following error. EINVAL The argument attr is invalid. f_pthread_attr_getschedparam(attr, param) Purpose
This function can be used to query the scheduling property setting in the thread attribute object attr. The current setting will be returned in the argument param. Class
Function Argument Type and Attributes attr TYPE(f_pthread_attr_t), INTENT(IN) param TYPE(f_sched_param), INTENT(OUT) Result Type and Attributes
INTEGER(4) Result Value
On successful completion, this function returns 0. Otherwise, this function returns the following error.
Chapter 7. Parallel programming with XL Fortran 207 EINVAL The argument attr is invalid. f_pthread_attr_getschedpolicy(attr, policy) Purpose
This function can be used to query the scheduling policy attribute setting in the attribute object attr. The current setting of the scheduling policy will be returned in the argument policy. Class
Function Argument Type and Attributes attr TYPE(f_pthread_attr_t), INTENT(IN) policy INTEGER(4), INTENT(OUT) Result Type and Attributes
INTEGER(4) Result Value
On successful completion, this function returns 0. Otherwise, this function returns the following error. EINVAL The argument attr is invalid. f_pthread_attr_getscope(attr, scope) Purpose
This function can be used to query the current setting of the scheduling scope attribute in the thread attribute object attr. The current setting will be returned through the argument scope. Class
Function Argument Type and Attributes attr TYPE(f_pthread_attr_t), INTENT(IN) scope INTEGER(4), INTENT(OUT) On return from the function, scope will contain one of the following values: PTHREAD_SCOPE_SYSTEM: the thread will compete for system resources on a system wide scope. PTHREAD_SCOPE_PROCESS: the thread will compete for system resources locally within the owning process.
208 XL Fortran: Optimization and Programming Guide scope Contains the following value: PTHREAD_SCOPE_SYSTEM: the thread will compete for system resources on a system wide scope. Result Type and Attributes
INTEGER(4) Result Value
On successful completion, this function returns 0. Otherwise, this function returns the following error. EINVAL The argument attr is invalid. f_pthread_attr_getstack(attr, stackaddr, ssize) Purpose
Retrieves the values of the stackaddr and stacksize arguments from the thread attribute object attr. Class
Function Argument Type and Attributes attr TYPE(f_pthread_attr_t), INTENT(IN) stackaddr Integer pointer, INTENT(OUT) ssize INTEGER(KIND=register_size), INTENT(OUT) where register_size is 8 in 64-bit mode. Result Type and Attributes
INTEGER(4) Result Value
On successful completion, this function returns 0. Otherwise, this function returns the following error. EINVAL One or more of the supplied arguments is invalid. f_pthread_attr_init(attr) Purpose
This function must be called to create and initialize the pthread attribute object attr before it can be used in any way. It will be filled with system default thread attribute values. After it is initialized, certain pthread attributes can be changed and/or set through attribute access procedures. Once initialized, this attribute
Chapter 7. Parallel programming with XL Fortran 209 object can be used to create a thread with the intended attributes. Class
Function Argument Type and Attributes attr TYPE(f_pthread_attr_t), INTENT(OUT) Result Type and Attributes
INTEGER(4) Result Value
On successful completion, this function returns 0. Otherwise, this function returns the following error. EINVAL The argument attr is invalid. f_pthread_attr_setdetachstate(attr, detach) Purpose
This function can be used to set the detach state attribute in the thread attribute object attr. Class
Function Argument Type and Attributes attr TYPE(f_pthread_attr_t), INTENT(OUT) detach INTEGER(4), INTENT(IN) Must contain one of the following values: PTHREAD_CREATE_DETACHED: when a thread attribute object of this attribute setting is used to create a new thread, the newly created thread will be in detached state. This is the system default setting. PTHREAD_CREATE_JOINABLE: when a thread attribute object of this attribute setting is used to create a new thread, the newly created thread will be in undetached state. Result Type and Attributes
INTEGER(4) Result Value
On successful completion, this function returns 0. Otherwise, this function returns the following error.
210 XL Fortran: Optimization and Programming Guide EINVAL The argument detach is invalid. f_pthread_attr_setguardsize(attr, guardsize) Purpose
This function is used to set the guardsizeattribute in the thread attributes object attr. The new value of this attribute is obtained from the argument guardsize.If guardsize is zero, a guard area will not be provided for threads created with attr. If guardsize is greater than zero, a guard area of at least sizeguardsize bytes is provided for each thread created with attr. Class
Function Argument Type and Attributes attr TYPE(f_pthread_attr_t), INTENT(INOUT) guardsize INTEGER(KIND=register_size), INTENT(IN) where register_size is 8 in 64-bit mode. Result Type and Attributes
INTEGER(4) Result Value
On successful completion, this function returns 0. Otherwise, this function returns the following error. EINVAL The argument attr or the argument guardsize is invalid. f_pthread_attr_setinheritsched(attr, inherit) Purpose
This function can be used to set the inheritance attribute of the thread scheduling property in the thread attribute object attr. Class
Function Argument Type and Attributes attr TYPE(f_pthread_attr_t), INTENT(OUT) inherit INTEGER(4), INTENT(IN) Must contain one of the following values: PTHREAD_INHERIT_SCHED: indicating that newly created threads will inherit the scheduling
Chapter 7. Parallel programming with XL Fortran 211 property of the parent thread and ignore the scheduling property of the thread attribute object used to create them. PTHREAD_EXPLICIT_SCHED: the scheduling property in the thread attribute object will be assigned to the newly created threads when it is used to create them. Result Type and Attributes
INTEGER(4) Result Value
On successful completion, this function returns 0. Otherwise, this function returns the following error. EINVAL The argument inherit is invalid. f_pthread_attr_setschedparam(attr, param) Purpose
This function can be used to set the scheduling property attribute in the thread attribute object attr. Threads created with this new attribute object will assume the scheduling property of argument param if they are not inherited from the creating thread. The sched_priority field in argument param indicates the thread's scheduling priority. The priority field must assume a value in the range of 1-127, where 127 is the most favored scheduling priority while 1 is the least. Class
Function Argument Type and Attributes attr TYPE(f_pthread_attr_t), INTENT(INOUT) param TYPE(f_sched_param), INTENT(IN) Result Type and Attributes
INTEGER(4) Result Value
On successful completion, this function returns 0. Otherwise, this function returns the following error. EINVAL The argument param is invalid.
212 XL Fortran: Optimization and Programming Guide f_pthread_attr_setschedpolicy(attr, policy) Purpose
After the attribute object is set by this function, threads created with this attribute object will assume the set scheduling policy if the scheduling property is not inherited from the creating thread. Class
Function Argument Type and Attributes attr TYPE(f_pthread_attr_t), INTENT(INOUT) policy INTEGER(4), INTENT(IN) Must contain one of the following values: SCHED_FIFO: indicating a first-in first-out thread scheduling policy. SCHED_RR: indicating a round-robin scheduling policy. SCHED_OTHER: the default scheduling policy. Result Type and Attributes
INTEGER(4) Result Value
On successful completion, this function returns 0. Otherwise, this function returns one of the following error. EINVAL The argument policy is invalid. f_pthread_attr_setscope(attr, scope) Purpose
This function can be used to set the contention scope attribute in the thread attribute object attr. Class
Function Argument Type and Attributes attr TYPE(f_pthread_attr_t), INTENT(INOUT) scope INTEGER(4), INTENT(IN) Must contain one of the following values:
Chapter 7. Parallel programming with XL Fortran 213 PTHREAD_SCOPE_SYSTEM: the thread will compete for system resources on a system wide scope. PTHREAD_SCOPE_PROCESS: the thread will compete for system resources locally within the owning process. Result Type and Attributes
INTEGER(4) Result Value
On successful completion, this function returns 0. Otherwise, this function returns the following error. EINVAL The argument scope is invalid. f_pthread_attr_setstack(attr, stackaddr, ssize) Purpose
Use this function to set the stack address and stack size attributes in the pthread attribute object attr. The stackaddr argument represents the stack address as an Integer pointer. The stacksize argument is an integer that represents the size of the stack in bytes. When creating a thread using the attribute object attr, the system allocates a minimum stack size of stacksize bytes. Class
Function Argument Type and Attributes attr TYPE(f_pthread_attr_t), INTENT(INOUT) stackaddr Integer pointer, INTENT(IN) ssize INTEGER(KIND=register_size) where register_size is 8 in 64-bit mode. Result Type and Attributes
INTEGER(4) Result Value
On successful completion, this function returns 0. Otherwise, this function returns one of the following errors. EINVAL The value of one or both of the supplied arguments is invalid. EACCES The stack pages specified are not readable by the thread.
214 XL Fortran: Optimization and Programming Guide f_pthread_attr_t Purpose
A derived data type whose components are all private. Any object of this type should be manipulated only through the appropriate interfaces provided in this module.
This data type corresponds to the POSIX pthread_attr_t, which is the type of thread attribute object. Class
Data Type. f_pthread_cancel(thread) Purpose
This function can be used to cancel a target thread. How this cancelation request will be processed depends on the state of the cancelability of the target thread. The target thread is identified by argument thread. If the target thread is in deferred-cancel state, this cancelation request will be put on hold until the target thread reaches its next cancelation point. If the target thread disables its cancelability, this request will be put on hold until it is enabled again. If the target thread is in async-cancel state, this request will be acted upon immediately. Class
Function Argument Type and Attributes thread TYPE(f_pthread_t), INTENT(INOUT) Result Type and Attributes
INTEGER(4) Result Value
On successful completion, this function returns 0. Otherwise, this function returns the following error. ESRCH The argument thread is invalid. f_pthread_cleanup_pop(exec) Purpose
This subroutine should be paired with f_pthread_cleanup_push in using the cleanup stack for thread safety. If the supplied argument exec contains a non-zero value, the last pushed cleanup function will be popped from the cleanup stack and executed, with the argument arg (from the last f_pthread_cleanup_push) passed to the cleanup function.
Chapter 7. Parallel programming with XL Fortran 215 If exec contains a zero value, the last pushed cleanup function will be popped from the cleanup stack, but will not be executed. Class
Subroutine Argument Type and Attributes exec INTEGER(4), INTENT(IN) Result Type and Attributes
None. Result Value
None. f_pthread_cleanup_push(cleanup, flag, arg) Purpose
This function can be used to register a cleanup subroutine for the calling thread. In case of an unexpected termination of the calling thread, the system will automatically execute the cleanup subroutine in order for the calling thread to terminate safely. The argument cleanup must be a subroutine expecting exactly one argument. If it is executed, the argument arg will be passed to it as the actual argument.
The argument arg is a generic argument that can be of any type and any rank. The actual argument arg must be a variable, and consequently eligible as a left-value in an assignment statement. If you pass an array section with vector subscripts to the argument arg, the result is unpredictable.
If the actual argument arg is an array section, the corresponding dummy argument in subroutine cleanup must be an assumed-shape array. Otherwise, the result is unpredictable.
If the actual argument arg has the pointer attribute that points to an array or array section, the corresponding dummy argument in subroutine cleanup must have a pointer attribute or be an assumed-shape array. Otherwise, the result is unpredictable.
For a normal execution path, this function must be paired with a call to f_pthread_cleanup_pop.
The argument flag must be used to convey the property of argument arg exactly to the system. Class
Function
216 XL Fortran: Optimization and Programming Guide Argument Type and Attributes cleanup A subroutine that has one dummy argument. flag Flag is an INTEGER(4), INTENT(IN) argument that can contain one of, or a combination of, the following constants: FLAG_CHARACTER: if the entry subroutine cleanup expects an argument of type CHARACTER in any way or any form, this flag value must be included to indicate this fact. However, if the subroutine expects a Fortran 90 pointer pointing to an argument of type CHARACTER, the FLAG_DEFAULT value should be included instead. FLAG_ASSUMED_SHAPE: if the entry subroutine cleanup has a dummy argument that is an assumed-shape array of any rank, this flag value must be included to indicate this fact. FLAG_DEFAULT: otherwise, this flag value is needed. arg A generic argument that can be of any type, kind, and rank. Result Type and Attributes
INTEGER(4) Result Value
On successful completion, this function returns 0. Otherwise, this function returns one of the following errors. ENOMEM The system cannot allocate memory to push this routine. EAGAIN The system cannot allocate resources to push this routine. EINVAL The argument flag is invalid. f_pthread_cond_broadcast(cond) Purpose
This function will unblock all threads waiting on the condition variable cond.If there is no thread waiting on this condition variable, the function will still succeed, but the next caller to f_pthread_cond_wait will be blocked, and will wait on the condition variable cond. Class
Function Argument Type and Attributes cond TYPE(f_pthread_cond_t), INTENT(INOUT)
Chapter 7. Parallel programming with XL Fortran 217 Result Type and Attributes
INTEGER(4) Result Value
On successful completion, this function returns 0. Otherwise, this function returns following error. EINVAL The argument cond is invalid. f_pthread_cond_destroy(cond) Purpose
This function can be used to destroy those condition variables that are no longer required. The target condition variable is identified by the argument cond. System resources allocated during initialization will be recollected by the system. Class
Function Argument Type and Attributes cond TYPE(f_pthread_cond_t), INTENT(INOUT) Result Type and Attributes
INTEGER(4) Result Value
On successful completion, this function returns 0. Otherwise, this function returns the following error. EBUSY The condition variable cond is being used by another thread. f_pthread_cond_init(cond, cattr) Purpose
This function can be used to dynamically initialize a condition variable cond. Its attributes will be set according to the attribute object cattr, if it is provided; otherwise, its attributes will be set to the system default. After the condition variable is initialized successfully, it can be used to synchronize threads.
Another method of initializing a condition variable is to initialize it statically using the Fortran constant PTHREAD_COND_INITIALIZER. Class
Function Argument Type and Attributes cond TYPE(f_pthread_cond_t), INTENT(INOUT)
218 XL Fortran: Optimization and Programming Guide cattr TYPE(f_pthread_condattr_t), INTENT(IN), OPTIONAL Result Type and Attributes
INTEGER(4) Result Value
On successful completion, this function returns 0. Otherwise, this function returns one of the following errors. EBUSY The condition variable is already in use. It is initialized and not destroyed. EINVAL The argument cond or cattr is invalid. f_pthread_cond_signal(cond) Purpose
This function will unblock at least one thread waiting on the condition variable cond. If there is no thread waiting on this condition variable, the function will still succeed, but the next caller to f_pthread_cond_wait will be blocked, and will wait on the condition variable cond. Class
Function Argument Type and Attributes cond TYPE(f_pthread_cond_t), INTENT(INOUT) Result Type and Attributes
INTEGER(4) Result Value
On successful completion, this function returns 0. Otherwise, this function returns the following error. EINVAL The argument cond is invalid. f_pthread_cond_t Purpose
A derived data type whose components are all private. Any object of this type should be manipulated through the appropriate interfaces provided in this module. In addition, objects of this type can be initialized at compile time using the Fortran constant PTHREAD_COND_INITIALIZER.
This data type corresponds to the POSIX pthread_cond_t, which is the type of condition variable object.
Chapter 7. Parallel programming with XL Fortran 219 Class
Data Type. f_pthread_cond_timedwait(cond, mutex, timeout) Purpose
This function can be used to wait for a certain condition to occur. The argument mutex must be locked before calling this function. The mutex is unlocked atomically and the calling thread waits for the condition to occur. The argument timeout specifies a deadline before which the condition must occur. If the deadline is reached before the condition occurs, the function will return an error code. This function provides a cancelation point in that the calling thread can be canceled if it is in the enabled state.
The argument timeout will specify an absolute date of the form: Oct. 31 10:00:53, 1998. For related information, see f_maketime and f_timespec. Class
Function Argument Type and Attributes cond TYPE(f_pthread_cond_t), INTENT(INOUT) mutex TYPE(f_pthread_mutex_t), INTENT(INOUT) timeout TYPE(f_timespec), INTENT(IN) Result Type and Attributes
INTEGER(4) Result Value
On successful completion, this function returns 0. Otherwise this function returns one of the following errors: EINVAL The argument cond, mutex,ortimeout is invalid. ETIMEDOUT The waiting deadline was reached before the condition occurred. f_pthread_cond_wait(cond, mutex) Purpose
This function can be used to wait for a certain condition to occur. The argument mutex must be locked before calling this function. The mutex is unlocked atomically, and the calling thread waits for the condition to occur. If the condition does not occur, the function will wait until the calling thread is terminated in another way. This function provides a cancelation point in that the calling thread can be canceled if it is in the enabled state.
220 XL Fortran: Optimization and Programming Guide Class
Function Argument Type and Attributes cond TYPE(f_pthread_cond_t), INTENT(INOUT) mutex TYPE(f_pthread_mutex_t), INTENT(INOUT) Result Type and Attributes
INTEGER(4) Result Value
This function returns 0. f_pthread_condattr_destroy(cattr) Purpose
This function can be called to destroy the condition variable attribute objects that are no longer required. The target object is identified by the argument cattr. The system resources allocated when it is initialized will be recollected. Class
Function Argument Type and Attributes cattr TYPE(f_pthread_condattr_t), INTENT(INOUT) Result Type and Attributes
INTEGER(4) Result Value
On successful completion, this function returns 0. Otherwise, this function returns on of the following errors. EINVAL The argument cattr is invalid. EBUSY Returns EBUSY if threads are waiting on the for the condition to occur. f_pthread_condattr_getpshared(cattr, pshared) Purpose
This function can be used to query the process-shared attribute of the condition variable attributes object identified by the argument cattr. The current setting of this attribute will be returned in the argument pshared.
Chapter 7. Parallel programming with XL Fortran 221 Class
Function Argument Type and Attributes cattr TYPE(f_pthread_condattr_t), INTENT(IN) pshared INTEGER(4), INTENT(OUT) On successful completion, pshared contains one of the following values: PTHREAD_PROCESS_SHARED The condition variable can be used by any thread that has access to the memory where it is allocated, even if these threads belong to different processes. PTHREAD_PROCESS_PRIVATE The condition variable shall only be used by threads within the same process as the thread that created it. Result Type and Attributes
INTEGER(4) Result Value
On successful completion, this function returns 0. Otherwise, this function returns the following error. EINVAL The argument cattr is invalid. f_pthread_condattr_init(cattr) Purpose
Use this function to initialize a condition variable attributes object cattr with the default value for all of the attributes defined by the implementation. Attempting to initialize an already initialized condition variable attributes object results in undefined behavior. After a condition variable attributes object has been used to initialize one or more condition variables, any function affecting the attributes object (including destruction) does not affect any previously initialized condition variables. Class
Function Argument Type and Attributes cattr TYPE(f_pthread_condattr_t), INTENT(OUT) Result Type and Attributes
INTEGER(4)
222 XL Fortran: Optimization and Programming Guide Result Value
On successful completion, this function returns 0. Otherwise, this function returns the following error. ENOMEM There is insufficient memory to initialize the condition variable attributes object. f_pthread_condattr_setpshared(cattr, pshared) Purpose
This function is used to set the process-shared attribute of the condition variable attributes object identified by the argument cattr. Its process-shared attribute will be set according to the argument pshared. Class
Function Argument Type and Attributes cattr TYPE(f_pthread_condattr_t), INTENT(INOUT) pshared is an INTEGER(4), INTENT(IN) argument that must contain one of the following values: PTHREAD_PROCESS_SHARED Specifies that the condition variable can be used by any thread that has access to the memory where it is allocated, even if these threads belong to different processes. PTHREAD_PROCESS_PRIVATE Specifies that the condition variable shall only be used by threads within the same process as the thread that created it. This is the default setting of the attribute. Result Type and Attributes
INTEGER(4) Result Value
On successful completion, this function returns 0. Otherwise, this function returns the following error. EINVAL The value specified by the argument cattr or pshared is invalid. f_pthread_condattr_t Purpose
A derived data type whose components are all private. Any object of this type should be manipulated only through the appropriate interfaces provided in this module.
Chapter 7. Parallel programming with XL Fortran 223 This data type corresponds to the POSIX pthread_condattr_t, which is the type of condition variable attribute object. Class
Data Type f_pthread_create(thread, attr, flag, ent, arg) Purpose
This function is used to create a new thread in the current process. The newly created thread will assume the attributes defined in the thread attribute object attr, if it is provided. Otherwise, the new thread will have system default attributes. The new thread will begin execution at the subroutine ent, which is required to have one dummy argument. The system will pass the argument arg to the thread entry subroutine ent as its actual argument. The argument flag is used to inform the system of the property of the argument arg. When the execution returns from the entry subroutine ent, the new thread will terminate automatically.
If subroutine ent was declared such that an explicit interface would be required if it was called directly, then an explicit interface is also required when it is passed as an argument to this function.
The argument arg is a generic argument that can be of any type and any rank. The actual argument arg must be a variable, and consequently eligible as a left- value in an assignment statement. If you pass an array section with vector subscripts to the argument arg, the result is unpredictable.
If the actual argument arg is an array section, the corresponding dummy argument in subroutine ent must be an assumed-shape array. Otherwise, the result is unpredictable.
If the actual argument arg has the pointer attribute that points to an array or array section, the corresponding dummy argument in subroutine ent must have a pointer attribute or be an assumed-shape array. Otherwise, the result is unpredictable. Class
Function Argument Type and Attributes thread TYPE(f_pthread_t), INTENT(OUT) On successful completion of the function, f_pthread_create stores the ID of the created thread in the thread. attr TYPE(f_pthread_attr_t), INTENT(IN) flag INTEGER(4), INTENT(IN) The argument flag must convey the property of the argument arg exactly to the system. The argument flag can be one of, or a combination of, the following constants: FLAG_CHARACTER: if the entry subroutine ent expects an argument of type
224 XL Fortran: Optimization and Programming Guide CHARACTER in any way or any form, this flag value must be included to indicate this fact. However, if the subroutine expects a Fortran 90 pointer pointing to an argument of type CHARACTER, the FLAG_DEFAULT value should be included instead. FLAG_ASSUMED_SHAPE: if the entry subroutine ent has a dummy argument which is an assumed-shape array of any rank, this flag value must be included to indicate this fact. FLAG_DEFAULT: otherwise, this flag value is needed. ent A subroutine that has one dummy argument of any type, kind and rank. arg A generic argument of any type, kind, and rank. It is passed to ent as the only actual argument. Result Type and Attributes
INTEGER(4) Result Value
On successful completion, this function returns 0. Otherwise, this function returns one of the following errors. EAGAIN The system does not have enough resources to create a new thread. EINVAL The argument thread, attr,orflag is invalid. ENOMEM The system does not have sufficient memory to create a new thread. f_pthread_detach(thread) Purpose
This function is used to indicate to the pthreads library implementation that storage for the thread whose thread ID is specified by the argument thread can be claimed when this thread terminates. If the thread has not yet terminated, f_pthread_detach shall not cause it to terminate. Multiple f_pthread_detach calls on the same target thread cause an error. Class
Function Argument Type and Attributes thread TYPE(f_pthread_t), INTENT(IN) Result Type and Attributes
INTEGER(4)
Chapter 7. Parallel programming with XL Fortran 225 Result Value
On successful completion, this function returns 0. Otherwise, this function returns the following error. ESRCH The argument thread is invalid. f_pthread_equal(thread1, thread2) Purpose
This function can be used to compare whether two thread ID's identify the same thread or not. Class
Function Argument Type and Attributes thread1 TYPE(f_pthread_t), INTENT(IN) thread2 TYPE(f_pthread_t), INTENT(IN) Result Type and Attributes
LOGICAL(4) Result Value TRUE The two thread ID's identify the same thread. FALSE The two thread ID's do not identify the same thread. f_pthread_exit(ret) Purpose
This subroutine can be called explicitly to terminate the calling thread before it returns from the entry subroutine. The actions taken depend on the state of the calling thread. If it is in non-detached state, the calling thread will wait to be joined. If the thread is in detached state, or when it is joined by another thread, the calling thread will terminate safely. First, the cleanup stack will be popped and executed, and then any thread-specific data will be destructed by the destructors. Finally, the thread resources are freed and the argument ret will be returned to the joining threads. The argument ret of this subroutine is optional. Currently, argument ret is limited to be an Integer pointer. If it is not an Integer pointer, the behavior is undefined. Calling f_pthread_exit will not automatically free all of the memory allocated to a thread. To avoid memory leaks, finalization must be handled separately from f_pthread_exit.
This subroutine never returns. If argument ret is not provided, NULL will be provided as this thread's exit status.
226 XL Fortran: Optimization and Programming Guide Class
Subroutine Argument Type and Attributes ret Integer pointer, OPTIONAL, INTENT(IN) Result Type and Attributes
None Result Value
None f_pthread_getconcurrency() Purpose
This function returns the value of the concurrency level set by a previous call to the f_pthread_setconcurrency function. If the f_pthread_setconcurrency function was not previously called, this function returns zero to indicate that the system is maintaining the concurrency level. Class
Function Argument Type and Attributes
None Result Type and Attributes
INTEGER(4) Result Value
This function returns the value of the concurrency level set by a previous call to the f_pthread_setconcurrency function. If the f_pthread_setconcurrency function was not previously called, this function returns 0. f_pthread_getschedparam(thread, policy, param) Purpose
This function can be used to query the current setting of the scheduling property of the target thread. The target thread is identified by argument thread. Its scheduling policy will be returned through argument policy and its scheduling property through argument param. The sched_priority field in param defines the scheduling priority. The priority field will assume a value in the range of 1-127, where 127 is the most favored scheduling priority while 1 is the least. Class
Function
Chapter 7. Parallel programming with XL Fortran 227 Argument Type and Attributes thread TYPE(f_pthread_t), INTENT(IN) policy INTEGER(4), INTENT(OUT) param TYPE(f_sched_param), INTENT(OUT) Result Type and Attributes
INTEGER(4) Result Value
On successful completion, this function returns 0. Otherwise, this function returns one of the following errors. ESRCH The target thread is invalid or has already terminated. EFAULT The policy or param points are outside the process memory space f_pthread_getspecific(key, arg) Purpose
This function can be used to retrieve the thread-specific data associated with key. Note that the argument arg is not optional in this function as it will return the thread-specific data. After execution of the procedure, the argument arg holds a pointer to the data, or NULL if there is no data to retrieve. The argument arg must be an Integer pointer, or the result is undefined.
The actual argument arg must be a variable, and consequently eligible as a left-value in an assignment statement. If you pass an array section with vector subscripts to the argument arg, the result is unpredictable. Class
Function Argument Type and Attributes key TYPE(f_pthread_key_t), INTENT(IN) arg Integer pointer, INTENT(OUT) Result Type and Attributes
INTEGER(4) Result Value
On successful completion, this function returns 0. Otherwise, this function returns the following error. EINVAL The argument key is invalid.
228 XL Fortran: Optimization and Programming Guide f_pthread_join(thread, ret) Purpose
This function can be called to join a particular thread designated by the argument thread. If the target thread is in non-detached state and is already terminated, this call will return immediately with the target thread's status returned in argument ret if it is provided. The argument ret is optional. Currently, ret must be an Integer pointer if it is provided.
If the target thread is in detached state, it is an error to join it. Class
Function Argument Type and Attributes thread TYPE(f_pthread_t), INTENT(IN) ret Integer pointer, INTENT(OUT), OPTIONAL Result Type and Attributes
INTEGER(4) Result Value
On successful completion, this function returns 0. Otherwise, this function returns one of the following errors. EDEADLK This call will cause a deadlock, or the calling thread is trying to join itself. EINVAL The argument thread is invalid. ESRCH The argument thread designates a thread which does not exist or is in detached state. f_pthread_key_create(key, dtr) Purpose
This function can be used to acquire a thread-specific data key. The key will be returned in the argument key. The argument dtr is a subroutine that will be used to destruct the thread-specific data associated with this key when any thread terminates after this calling point. The destructor will receive the thread-specific data as its argument. The destructor itself is optional. If it is not provided, the system will not invoke any destructor on the thread-specific data associated with this key. Note that the number of thread-specific data keys is limited in each process. It is the user's responsibility to manage the usage of the keys. The per-process limit can be checked by the Fortran constant PTHREAD_DATAKEYS_MAX. Class
Function
Chapter 7. Parallel programming with XL Fortran 229 Argument Type and Attributes key TYPE(f_pthread_key_t), INTENT(OUT) dtr External, optional subroutine Result Type and Attributes
INTEGER(4) Result Value
On successful completion, this function returns 0. Otherwise, this function returns one of the following errors. EAGAIN The maximum number of keys has been exceeded. EINVAL The argument key is invalid. ENOMEM There is insufficient memory to create this key. f_pthread_key_delete(key) Purpose
This function will destroy the thread-specific data key identified by the argument key. It is the user's responsibility to ensure that there is no thread-specific data associated with this key. This function does not call any destructor on the thread's behalf. After the key is destroyed, it can be reused by the system for f_pthread_key_create requests. Class
Function Argument Type and Attributes key TYPE(f_pthread_key_t), INTENT(INOUT) Result Type and Attributes
INTEGER(4) Result Value
On successful completion, this function returns 0. Otherwise, this function returns one of the following errors. EINVAL The argument key is invalid. EBUSY There is still data associated with this key.
230 XL Fortran: Optimization and Programming Guide f_pthread_key_t Purpose
A derived data type whose components are all private. Any object of this type should be manipulated only through the appropriate interfaces provided in this module.
This data type corresponds to the POSIX pthread_key_t, which is the type of key object for accessing thread-specific data. Class
Data Type f_pthread_kill(thread, sig) Purpose
This function can be used to send a signal to a target thread. The target thread is identified by argument thread. The signal which will be sent to the target thread is identified in argument sig.Ifsig contains value zero, error checking will be done by the system but no signal will be sent. Class
Function Argument Type and Attributes thread TYPE(f_pthread_t), INTENT(INOUT) sig INTEGER(4), INTENT(IN) Result Type and Attributes
INTEGER(4) Result Value
On successful completion, this function returns 0. Otherwise, this function returns one of the following errors. EINVAL The argument thread or sig is invalid. ESRCH The target thread does not exist. f_pthread_mutex_destroy(mutex) Purpose
This function should be called to destroy those mutex objects that are no longer required. In this way, the system can recollect the memory resources. The target mutex object is identified by the argument mutex.
Chapter 7. Parallel programming with XL Fortran 231 Class
Function Argument Type and Attributes mutex TYPE(f_pthread_mutex_t), INTENT(INOUT) Result Type and Attributes
INTEGER(4) Result Value
On successful completion, this function returns 0. Otherwise, this function returns one of the following errors. EBUSY The target mutex is locked or referenced by another thread. EINVAL The argument mutex is invalid. f_pthread_mutex_init(mutex, mattr) Purpose
This function can be used to initialize the mutex object identified by argument mutex. The initialized mutex will assume attributes set in the mutex attribute object mattr, if it is provided. If mattr is not provided, the system will initialize the mutex to have default attributes. After it is initialized, the mutex object can be used to synchronize accesses to critical data or code. It can also be used to build more complicated thread synchronization objects.
Another method to initialize mutex objects is to statically initialize them through the Fortran constant PTHREAD_MUTEX_INITIALIZER. If this method of initialization is used it is not necessary to call the function before using the mutex objects. Class
Function Argument Type and Attributes mutex TYPE(f_pthread_mutex_t), INTENT(OUT) mattr TYPE(f_pthread_mutexattr_t), INTENT(IN), OPTIONAL Result Type and Attributes
INTEGER(4) Result Value
This function always returns 0.
232 XL Fortran: Optimization and Programming Guide f_pthread_mutex_lock(mutex) Purpose
This function can be used to acquire ownership of the mutex object. (In other words, the function will lock the mutex.) If the mutex has already been locked by another thread, the caller will wait until the mutex is unlocked. If the mutex is already locked by the caller itself, an error will be returned to prevent recursive locking. Class
Function Argument Type and Attributes mutex TYPE(f_pthread_mutex_t), INTENT(INOUT) Result Type and Attributes
INTEGER(4) Result Value
On successful completion, this function returns 0. Otherwise, this function returns one of the following errors. EDEADLK The mutex is locked by the calling thread already. EINVAL The argument mutex is invalid. f_pthread_mutex_t Purpose
A derived data type whose components are all private. Any object of this type should be manipulated through the appropriate interfaces provided in this module. In addition, objects of this type can be initialized statically through the Fortran constant PTHREAD_MUTEX_INITIALIZER.
This data type corresponds to the POSIX pthread_mutex_t, which is the type of mutex object. Class
Data Type f_pthread_mutex_trylock(mutex) Purpose
This function can be used to acquire ownership of the mutex object. (In other words, the function will lock the mutex.) If the mutex has already been locked by another thread, the function returns the error code EBUSY. The calling thread can check the return code to take further actions. If the mutex is already locked by the caller itself, an error will be returned to prevent recursive locking.
Chapter 7. Parallel programming with XL Fortran 233 Class
Function Argument Type and Attributes mutex TYPE(f_pthread_mutex_t), INTENT(INOUT) Result Type and Attributes
INTEGER(4) Result Value
On successful completion, this function returns 0. Otherwise, this function returns one of the following errors. EBUSY The target mutex is locked or referenced by another thread. EINVAL The argument mutex is invalid. f_pthread_mutex_unlock(mutex) Purpose
This function releases the mutex object's ownership in order to allow other threads to lock the mutex. Class
Function Argument Type and Attributes mutex TYPE(f_pthread_mutex_t), INTENT(INOUT) Result Type and Attributes
INTEGER(4) Result Value
On successful completion, this function returns 0. Otherwise, this function returns one of the following errors. EINVAL The argument mutex is invalid. EPERM The mutex is not locked by the calling thread. f_pthread_mutexattr_destroy(mattr) Purpose
This function can be used to destroy a mutex attribute object that has been initialized previously. Allocated memory will then be recollected. A mutex created with this attribute will not be affected by this action.
234 XL Fortran: Optimization and Programming Guide Class
Function Argument Type and Attributes mattr TYPE(f_pthread_mutexattr_t), INTENT(INOUT) Result Type and Attributes
INTEGER(4) Result Value
This function always returns 0. f_pthread_mutexattr_getpshared(mattr, pshared) Purpose
This function is used to query the process-shared attribute in the mutex attributes object identified by the argument mattr. The current setting of the attribute will be returned through the argument pshared. Class
Function Argument Type and Attributes mattr TYPE(f_pthread_mutexattr_t), INTENT(IN) pshared INTEGER(4), INTENT(IN) On return from this function, pshared contains one of the following values: PTHREAD_PROCESS_SHARED The mutex can be operated upon by any thread that has access to the memory where the mutex is allocated, even if the mutex is allocated in memory that is shared by multiple processes. PTHREAD_PROCESS_PRIVATE The mutex will only be operated upon by threads created within the same process as the thread that initialized the mutex. Result Type and Attributes
INTEGER(4) Result Value
If this function completes successfully, value 0 is returned and the value of the process-shared attribute is returned through the argument pshared. Otherwise, the following error will be returned: EINVAL The argument mattr is invalid.
Chapter 7. Parallel programming with XL Fortran 235 f_pthread_mutexattr_gettype(mattr, type) Purpose
This function is used to query the mutex type attribute in the mutex attributes object identified by the argument mattr.
If this function completes successfully, value 0 is returned and the type attribute will be returned through the argument type. Class
Function Argument Type and Attributes mattr TYPE(f_pthread_mutexattr_t), INTENT(IN) type INTEGER(4), INTENT(OUT) On return from this function, type contains one of the following values: PTHREAD_MUTEX_NORMAL This type of mutex does not detect deadlock. A thread attempting to relock this mutex without first unlocking it will deadlock. Attempting to unlock a mutex locked by a different thread results in undefined behavior. PTHREAD_MUTEX_ERRORCHECK This type of mutex provides error checking. A thread attempting to relock this mutex without first unlocking it will return with an error. A thread attempting to unlock a mutex which another thread has locked will return an error. A thread attempting to unlock an unlocked mutex will return with an error. PTHREAD_MUTEX_RECURSIVE A thread attempting to relock this mutex without first unlocking it will succeed in locking the mutex. The relocking deadlock that can occur with mutexes of type PTHREAD_MUTEX_NORMAL cannot occur with this type of mutex. Multiple locks of this mutex require the same number of unlocks to release the mutex before another thread can acquire the mutex. Result Type and Attributes
INTEGER(4) Result Value
On successful completion, this function returns 0. Otherwise, this function returns the following error. EINVAL The argument is invalid.
236 XL Fortran: Optimization and Programming Guide f_pthread_mutexattr_init(mattr) Purpose
This function can be used to initialize a mutex attribute object before it can be used in any other way. The mutex attribute object will be returned through argument mattr. Class
Function Argument Type and Attributes mattr TYPE(f_pthread_mutexattr_t), INTENT(OUT) Result Type and Attributes
INTEGER(4) Result Value
This function returns 0. f_pthread_mutexattr_setpshared(mattr, pshared) Purpose
This function is used to set the process-shared attribute of the mutex attributes object identified by the argument mattr. Class
Function Argument Type and Attributes mattr TYPE(f_pthread_mutexattr_t), INTENT(INOUT) pshared INTEGER(4), INTENT(IN) Must contain one of the following values: PTHREAD_PROCESS_SHARED Specifies the mutex can be operated upon by any thread that has access to the memory where the mutex is allocated, even if the mutex is allocated in memory that is shared by multiple processes. PTHREAD_PROCESS_PRIVATE Specifies the mutex will only be operated upon by threads created within the same process as the thread that initialized the mutex. This is the default setting of the attribute. Result Type and Attributes
INTEGER(4)
Chapter 7. Parallel programming with XL Fortran 237 Result Value
On successful completion, this function returns 0. Otherwise, this function returns the following error. EINVAL The argument is invalid. f_pthread_mutexattr_settype(mattr, type) Purpose
This function is used to set the mutex type attribute in the mutex attributes object identified by the argument mattr The argument type identifies the mutex type attribute to be set. Class
Function Argument Type and Attributes mattr TYPE(f_pthread_mutexattr_t), INTENT(INOUT) type INTEGER(4), INTENT(IN) Must contain one of the following values: PTHREAD_MUTEX_NORMAL This type of mutex does not detect deadlock. A thread attempting to relock this mutex without first unlocking it will deadlock. Attempting to unlock a mutex locked by a different thread results in undefined behavior. PTHREAD_MUTEX_ERRORCHECK This type of mutex provides error checking. A thread attempting to relock this mutex without first unlocking it will return with an error. A thread attempting to unlock a mutex which another thread has locked will return an error. A thread attempting to unlock an unlocked mutex will return with an error. PTHREAD_MUTEX_RECURSIVE A thread attempting to relock this mutex without first unlocking it will succeed in locking the mutex. The relocking deadlock that can occur with mutexes of type PTHREAD_MUTEX_NORMAL cannot occur with this type of mutex. Multiple locks of this mutex require the same number of unlocks to release the mutex before another thread can acquire the mutex. PTHREAD_MUTEX_DEFAULT The same as PTHREAD_MUTEX_NORMAL. Result Type and Attributes
INTEGER(4) Result Value
On successful completion, this function returns 0. Otherwise, this function returns the following error.
238 XL Fortran: Optimization and Programming Guide EINVAL One of the arguments is invalid. f_pthread_mutexattr_t Purpose
A derived data type whose components are all private. Any object of this type should be manipulated only through the appropriate interfaces provided in this module.
This data type corresponds to the POSIX pthread_mutexattr_t, which is the type of mutex attribute object. Class
Data Type f_pthread_once(once, initr) Purpose
This function can be used to initialize those data required to be initialized only once. The first thread calling this function will call initr to do the initialization. Other threads calling this function afterwards will have no effect. Argument initr must be a subroutine without dummy arguments. Class
Function Argument Type and Attributes once TYPE(f_pthread_once_t), INTENT(INOUT) initr A subroutine that has no dummy arguments. Result Type and Attributes
INTEGER(4) Result Value
This function returns 0. f_pthread_once_t Purpose
A derived data type whose components are all private. Any object of this type should be manipulated through the appropriate interfaces provided in this module. However, objects of this type can only be initialized through the Fortran constant PTHREAD_ONCE_INIT. Class
Data Type
Chapter 7. Parallel programming with XL Fortran 239 f_pthread_rwlock_destroy(rwlock) Purpose
This function destroys the read-write lock object specified by the argument rwlock and releases any resources used by the lock. Class
Function Argument Type and Attributes rwlock TYPE(f_pthread_rwlock_t), INTENT(INOUT) Result Type and Attributes
INTEGER(4) Result Value
On successful completion, this function returns 0. Otherwise, this function returns one of the following errors. EBUSY The target read-write object is locked. f_pthread_rwlock_init(rwlock, rwattr) Purpose
This function initializes the read-write lock object specified by rwlock with the attribute specified by the argument rwattr. If the optional argument rwattr is not provided, the system will initialize the read-write lock object with the default attributes. After it is initialized, the lock can be used to synchronize access to critical data. With a read-write lock, many threads can have simultaneous read-only access to data, while only one thread can have write access at any given time and no other readers or writers are allowed.
Another method to initialize read-write lock objects is to statically initialize them through the Fortran constant PTHREAD_RWLOCK_INITIALIZER. If this method of initialization is used, it is not necessary to call this function before using the read-write lock objects. Class
Function Argument Type and Attributes rwlock TYPE(f_pthread_rwlock_t), INTENT(OUT) rwattr TYPE(f_pthread_rwlockattr_t), INTENT(IN), OPTIONAL Result Type and Attributes
INTEGER(4)
240 XL Fortran: Optimization and Programming Guide Result Value
This function returns 0. f_pthread_rwlock_rdlock(rwlock) Purpose
This function applies a read lock to the read-write lock specified by the argument rwlock. The calling thread acquires the read lock if a writer does not hold the lock and there are no writes blocked on the lock. Otherwise, the calling thread will not acquire the read lock. If the read lock is not acquired, the calling thread blocks (that is, it does not return from the f_pthread_rwlock_rdlock call) until it can acquire the lock. Results are undefined if the calling thread holds a write lock on rwlock at the time the call is made. A thread may hold multiple concurrent read locks on rwlock (that is, successfully call the f_pthread_rwlock_rdlock function n times). If so, the thread must perform matching unlocks (that is, it must call the f_pthread_rwlock_unlock function n times). Class
Function Argument Type and Attributes rwlock TYPE(f_pthread_rwlock_t), INTENT(INOUT) Result Type and Attributes
INTEGER(4) Result Value
On successful completion, this function returns 0. Otherwise, this function returns one of the following errors. EAGAIN The read-write lock could not be acquired because the maximum number of read locks for rwlock has been exceeded. EINVAL The argument rwlock does not refer to an initialized read-write lock object. f_pthread_rwlock_t Purpose
A derived data type whose components are all private. Any object of this type should be manipulated only through the appropriate interfaces provided in this module. In addition, objects of this type can be initialized statically through the Fortran constant PTHREAD_RWLOCK_INITIALIZER. Class
Data Type
Chapter 7. Parallel programming with XL Fortran 241 f_pthread_rwlock_tryrdlock(rwlock) Purpose
This function applies a read lock like the f_pthread_rwlock_rdlock function with the exception that the function fails if any thread holds a write lock on rwlock or there are writers blocked on rwlock. In that case, the function returns EBUSY. The calling thread can check the return code to take further actions. Class
Function Argument Type and Attributes rwlock TYPE(f_pthread_rwlock_t), INTENT(INOUT) Result Type and Attributes
INTEGER(4) Result Value
This function returns zero if the lock for reading on the read-write lock object specified by rwlock is acquired. Otherwise, the following error will be returned: EBUSY The read-write lock could not be acquired for reading because a writer holds the lock or was blocked on it. f_pthread_rwlock_trywrlock(rwlock) Purpose
This function applies a write lock like the f_pthread_rwlock_wrlock function with the exception that the function fails if any thread currently holds rwlock (for reading or writing). In that case, the function returns EBUSY. The calling thread can check the return code to take further actions. Class
Function Argument Type and Attributes rwlock TYPE(f_pthread_rwlock_t), INTENT(INOUT) Result Type and Attributes
INTEGER(4) Result Value
This function returns zero if the lock for writing on the read-write lock object specified by rwlock is acquired. Otherwise, the following error will be returned:
242 XL Fortran: Optimization and Programming Guide EBUSY The read-write lock could not be acquired for writing because it is already locked for reading or writing. f_pthread_rwlock_unlock(rwlock) Purpose
This function is used to release a lock held on the read-write lock object specified by the argument rwlock. If this function is called to release a read lock from the read-write lock object and there are other read locks currently held on this read-write lock object, the read-write lock object remains in the read locked state. If this function releases the calling thread's last read lock on this read-write lock object, then the calling thread is no longer one of the owners of the object. If this function releases the last read lock for this read-write lock object, the read-write lock object will be put in the unlocked state with no owners. Class
Function Argument Type and Attributes rwlock TYPE(f_pthread_rwlock_t), INTENT(INOUT) Result Type and Attributes
INTEGER(4) Result Value
On successful completion, this function returns 0. Otherwise, this function returns one of the following errors. EPERM The current thread does not own the read-write lock. f_pthread_rwlock_wrlock(rwlock) Purpose
This function applies a write lock to the read-write lock specified by the argument rwlock. The calling thread acquires the write lock if no other thread (reader or writer) holds the read-write lock rwlock. Otherwise, the thread blocks (that is, does not return from the f_pthread_rwlock_wrlock call) until it acquires the lock. Results are undefined if the calling thread holds the read-write lock (whether a read or write lock) at the time the call is made. Class
Function Argument Type and Attributes rwlock TYPE(f_pthread_rwlock_t), INTENT(INOUT)
Chapter 7. Parallel programming with XL Fortran 243 Result Type and Attributes
INTEGER(4) Result Value
On successful completion, this function returns 0. Otherwise, this function returns the following error. EINVAL The argument rwlock does not refer to an initialized read-write lock object. f_pthread_rwlockattr_destroy(rwattr) Purpose
This function destroys a read-write lock attributes object specified by the argument rwattr which has been initialized previously. A read-write lock created with this attribute will not be affected by the action. Class
Function Argument Type and Attributes rwattr TYPE(f_pthread_rwlockattr_t), INTENT(INOUT) Result Type and Attributes
INTEGER(4) Result Value
On successful completion, this function returns 0. Otherwise, this function returns the following error. EINVAL The argument rwattr is invalid. f_pthread_rwlockattr_getpshared(rwattr, pshared) Purpose
This function is used to obtain the value of the process-shared attribute from the initialized read-write lock attributes object specified by the argument rwattr. The current setting of this attribute will be returned in the argument pshared. Class
Function Argument Type and Attributes rwattr TYPE(f_pthread_rwlockattr_t), INTENT(IN) pshared INTEGER(4), INTENT(OUT)
244 XL Fortran: Optimization and Programming Guide On return from this function, the value of pshared will be one of the following: PTHREAD_PROCESS_SHARED The read-write lock can be operated upon by any thread that has access to the memory where it is allocated, even if these threads belong to different processes. PTHREAD_PROCESS_PRIVATE The read-write lock shall only be used by threads within the same process as the thread that created it. Result Type and Attributes
INTEGER(4) Result Value
If this function completes successfully, value 0 is returned and the value of the process-shared attribute of rwattr is stored into the object specified by the argument pshared. Otherwise, the following error will be returned: EINVAL The argument rwattr is invalid. f_pthread_rwlockattr_init(rwattr) Purpose
This function initializes a read-write lock attributes object specified by rwattr with the default value for all of the attributes. Class
Function Argument Type and Attributes rwattr TYPE(f_pthread_rwlockattr_t), INTENT(OUT) Result Type and Attributes
INTEGER(4) Result Value
On successful completion, this function returns 0. Otherwise, this function returns the following error. ENOMEM There is insufficient memory to initialize the read-write lock attributes object.
Chapter 7. Parallel programming with XL Fortran 245 f_pthread_rwlockattr_setpshared(rwattr, pshared) Purpose
This function is used to set the process-shared attribute in an initialized read-write lock attributes object specified by the argument rwattr, based on the value provided by the argument pshared. Class
Function Argument Type and Attributes rwattr TYPE(f_pthread_rwlockattr_t), INTENT(INOUT) pshared INTEGER(4), INTENT(IN) Must be one of the following: PTHREAD_PROCESS_SHARED Specifies the read-write lock can be operated upon by any thread that has access to the memory where it is allocated, even if these threads belong to different processes. PTHREAD_PROCESS_PRIVATE Specifies the read-write lock shall only be used by threads within the same process as the thread that created it. This is the default setting of the attribute. Result Type and Attributes
INTEGER(4) Result Value
On successful completion, this function returns 0. Otherwise, this function returns one of the following errors: EINVAL The argument rwattr is invalid. ENOSYS The value of pshared is equal to pthread_process_shared. f_pthread_rwlockattr_t Purpose
This is a derived data type whose components are all private. Any object of this type should be manipulated only through the appropriate interfaces provided in this module. Class
Data Type
246 XL Fortran: Optimization and Programming Guide f_pthread_self() Purpose
This function can be used to return the thread ID of the calling thread. Class
Function Argument Type and Attributes
None Result Type and Attributes
TYPE(f_pthread_t) Result Value
The calling thread's ID is returned. f_pthread_setcancelstate(state, oldstate) Purpose
This function can be used to set the thread's cancelability state. The new state will be set according to the argument state. The old state will be returned in the argument oldstate. Class
Function Argument Type and Attributes state INTEGER(4), INTENT(IN) Must contain one of the following: PTHREAD_CANCEL_DISABLE: The thread's cancelability is disabled. PTHREAD_CANCEL_ENABLE: The thread's cancelability is enabled. oldstate INTEGER(4), INTENT(OUT) On return from this function, oldstate will contain one of the following values: PTHREAD_CANCEL_DISABLE: The thread's cancelability is disabled. PTHREAD_CANCEL_ENABLE: The thread's cancelability is enabled. Result Type and Attributes
INTEGER(4)
Chapter 7. Parallel programming with XL Fortran 247 Result Value
On successful completion, this function returns 0. Otherwise, this function returns the following error. EINVAL The argument state is invalid. f_pthread_setcanceltype(type, oldtype) Purpose
This function can be used to set the thread's cancelability type. The new type will be set according to the argument type. The old type will be returned in argument oldtype. Class
Function Argument Type and Attributes type INTEGER(4), INTENT(IN) Must contain one of the following values: PTHREAD_CANCEL_DEFERRED: Cancelation request will be delayed until a cancelation point. PTHREAD_CANCEL_ASYNCHRONOUS: Cancelation request will be acted upon immediately. This may cause unexpected results. oldtype INTEGER(4), INTENT(OUT) On return from this procedure, oldtype will contain one of the following values: PTHREAD_CANCEL_DEFERRED: Cancelation request will be delayed until a cancelation point. PTHREAD_CANCEL_ASYNCHRONOUS: Cancelation request will be acted upon immediately. This may cause unexpected results. Result Type and Attributes
INTEGER(4) Result Value
On successful completion, this function returns 0. Otherwise, this function returns the following error. EINVAL The argument type is invalid.
248 XL Fortran: Optimization and Programming Guide f_pthread_setconcurrency(new_level) Purpose
This function is used to inform the pthreads library implementation of desired concurrency level as specified by the argument new_level. The actual level of concurrency provided by the implementation as a result of this function call is unspecified. Class
Function Argument Type and Attributes new_level INTEGER(4), INTENT(IN) Result Type and Attributes
INTEGER(4) Result Value
This function returns 0. f_pthread_setschedparam(thread, policy, param) Purpose
This function can be used to dynamically set the scheduling policy and the scheduling property of a thread. The target thread is identified by argument thread. The new scheduling policy for the target thread is provided through argument policy. The new scheduling property of the target thread will be set to the value provided by argument param. The sched_priority field in param defines the scheduling priority. Its range is 1-127. Class
Function Argument Type and Attributes thread TYPE(f_pthread_t), INTENT(INOUT) policy INTEGER(4), INTENT(IN) param TYPE(f_sched_param), INTENT(IN) Result Type and Attributes
INTEGER(4) Result Value
On successful completion, this function returns 0. Otherwise, this function returns one of the following errors
Chapter 7. Parallel programming with XL Fortran 249 ENOSYS The POSIX priority scheduling option is not implemented on Blue Gene/Q. ENOTSUP The value of argument policy or param is not supported. EPERM The target thread is not permitted to perform the operation or is in a mutex protocol already. ESRCH The target thread does not exist or is invalid. f_pthread_setspecific(key, arg) Purpose
This function can be used to set the calling thread's specific data associated with the key identified by argument key. The argument arg, which is optional, identifies the thread-specific data to be set. If arg is not provided, the thread-specific data will be set to NULL, which is the initial value for each thread. Only an Integer pointer can be passed as the arg argument. If arg is not an Integer pointer, the result is undefined.
The actual argument arg must be a variable, and consequently eligible as a left-value in an assignment statement. If you pass an array section with vector subscripts to the argument arg, the result is unpredictable. Class
Function Argument Type and Attributes key TYPE(f_pthread_key_t), INTENT(IN) arg Integer pointer, INTENT(IN), OPTIONAL Result Type and Attributes
INTEGER(4) Result Value
On successful completion, this function returns 0. Otherwise, this function returns one of the following errors EINVAL The argument key is invalid. ENOMEM There is insufficient memory to associate the data with the key.
250 XL Fortran: Optimization and Programming Guide f_pthread_t Purpose
A derived data type whose components are all private. Any object of this type should be manipulated only through the appropriate interfaces provided in this module.
This data type corresponds to the POSIX pthread_t, which is the type of thread object. Class
Data Type f_pthread_testcancel() Purpose
This subroutine provides a cancelation point in a thread. When it is called, any pending cancelation request will be acted upon immediately if it is in the enabled state. Class
Subroutine Argument Type and Attributes
None Result Type and Attributes
None f_sched_param Purpose
This data type corresponds to the Blue Gene/Q system data structure sched_param, which is a system data type.
This is a public data structure defined as: type f_sched_param sequence integer sched_priority end type f_sched_param
Class
Data Type f_sched_yield() Purpose
This function is used to force the calling thread to relinquish the processor until it again becomes the head of its thread list.
Chapter 7. Parallel programming with XL Fortran 251 Class
Function Argument Type and Attributes
None. Result Type and Attributes
INTEGER(4) Result Value
If this function completes successfully, value 0 is returned. Otherwise, a value of -1 will be returned. f_timespec Purpose
This is a Fortran definition of the Blue Gene/Q system data structure timespec. Within the Fortran Pthreads module, objects of this type are used to specify an absolute date and time. This deadline absolute date is used when waiting on a POSIX condition variable.
In 64–bit mode, f_timespec is defined as: TYPE F_Timespec SEQUENCE INTEGER(4) tv_sec INTEGER(4) pad INTEGER(KIND=REGISTER_SIZE) tv_nsec END TYPE F_Timespec
Class
Data Type
252 XL Fortran: Optimization and Programming Guide Chapter 8. Interlanguage calls
Your Fortran application can perform interlanguage calls to routines written in a language other than Fortran.
The guidelines assume that you are familiar with the syntax of all applicable languages.
Conventions for XL Fortran external names To assist you in writing mixed-language programs, XL Fortran follows a consistent set of rules when translating the name of a global entity into an external name that the linker can resolve.
The rules are: v Both the underscore (_) and the dollar sign ($) are valid characters anywhere in names. Because names that begin with an underscore are reserved for the names of library routines, do not use an underscore as the first character of a Fortran external name. To avoid conflicts between Fortran and non-Fortran function names, you can compile the Fortran program with the -qextname option. This option adds an underscore to the end of the Fortran names. Then use an underscore as the last character of any non-Fortran procedures that you want to call from Fortran. v Names can be up to 250 characters long. v Program and symbolic names are interpreted as all lowercase by default. If you are writing new non-Fortran code, use all-lowercase procedure names to simplify calling the procedures from Fortran. You can use the -U option or the @PROCESS MIXED directive if you want the names to use both uppercase and lowercase: @process mixed external C_Func ! With MIXED, we can call C_Func, not just c_func. integer aBc, ABC ! With MIXED, these are different variables. common /xYz/ aBc ! The same applies to the common block names. common /XYZ/ ABC ! xYz and XYZ are external names that are ! visible during linking. end v Names for module procedures are formed by concatenating __ (two underscores), the module name, _IMOD_ (for intrinsic modules) or _NMOD_ (for non-intrinsic modules), and the name of the module procedure. For example, module procedure MYPROC in module MYMOD has the external name __mymod_NMOD_myproc.
Note: Symbolic debuggers and other tools should account for this naming scheme when debugging XL Fortran programs that contain module procedures. For example, some debuggers default to lowercase for program and symbolic names. This behavior should be changed to use mixed case when debugging XL Fortran programs with module procedures. v The XL compilers generate code that uses main as an external entry point name. You can only use main as an external name in these contexts:
© Copyright IBM Corp. 1990, 2012 253 – A Fortran program or local-variable name. (This restriction means that you cannot use main as a binding label, or for the name of an external function, external subroutine, block data program unit, or common block. References to such an object use the compiler-generated main instead of your own.) – The name of the top-level main function inaCprogram. v Some other potential naming conflicts may occur when linking a program. For instructions on avoiding them, see Avoiding naming conflicts during linking in the XL Fortran Compiler Reference.
If you are porting your application from another system and your application does encounter naming conflicts like these, you may need to use the -qextname option.
Mixed-language input and output To improve performance, the XL Fortran runtime library has its own buffers and its own handling of these buffers. This means that mixed-language programs cannot freely mix I/O operations on the same file from the different languages.
Mixing code compiled by multiple Fortran compilers, for example xlf and gfortran, could face similar problems. The safest approach is to treat the code compiled by another Fortran compiler as non-Fortran code. To maintain data integrity in such cases: v If the file position is not important, open and explicitly close the file within the Fortran part of the program before performing any I/O operations on that file from subprograms written in another language. v To open a file in Fortran and manipulate the open file from another language, call the flush_ procedure to save any buffer for that file, and then use the getfd procedure to find the corresponding file descriptor and pass it to the non-Fortran subprogram. As an alternative to calling the flush_ procedure, you can use the buffering runtime option to disable the buffering for I/O operations. When you specify buffering=disable_preconn, XL Fortran disables the buffering for preconnected units. When you specify buffering=disable_all, XL Fortran disables the buffering for all logical units.
Note: After you call flush_ to flush the buffer for a file, do not do anything to the file from the Fortran part of the program except to close it when the non-Fortran processing is finished. v If any XL Fortran subprograms containing WRITE statements are called from a non-Fortran main program, explicitly CLOSE the data file, or use the flush_ subroutine in the XL Fortran subprograms to ensure that the buffers are flushed. Alternatively, you can use the buffering runtime option to disable buffering for I/O operations.
For more information on the flush_ and getfd procedures, see the Service and utility procedures topic in the XL Fortran Language Reference. For more information on the buffering runtime option, see Setting runtime options in the XL Fortran Compiler Reference.
Mixing Fortran and C++ When mixing Fortran and C++ in the same program, you need to invoke the C++ compiler to correctly link the final program.
254 XL Fortran: Optimization and Programming Guide Most of the information in this section applies to Fortran and languages with similar data types and naming schemes. However, to mix Fortran and C++ in the same program, you must add an extra level of indirection and pass the interlanguage calls through C++ wrapper functions.
Because the C++ compiler mangles the names of some C++ objects, you must use your C++ compiler's invocation command, like bgxlC or g++, to link the final program and include -L and -l options for the XL Fortran library directories and libraries.
program main
integer idim,idim1
idim = 35 idim1= 45
write(6,*) ’Inside Fortran calling first C function’ call cfun(idim) write(6,*) ’Inside Fortran calling second C function’ call cfun1(idim1) write(6,*) ’Exiting the Fortran program’ end
Figure 3. Main Fortran program that calls C++ (main1.f)
#include
extern "C" void cfun(int *idim){ printf("%%Inside C function before creating C++ Object\n"); int i = *idim; junk
extern "C" void cfun1(int *idim1) { printf("%%Inside C function cfun1 before creating C++ Object\n"); int i = *idim1; temp
Figure 4. C++ wrapper functions for calling C++ (cfun.C)
Chapter 8. Interlanguage calls 255 #include
template
private: int inter; T templ_mem; T stor_val;
public: junk(int i,T j): inter(i),templ_mem(j) {cout <<"***Inside C++ constructor" << endl;}
~junk() {cout <<"***Inside C++ Destructor" << endl;}
void store(T *val){ stor_val = *val;}
void print(void) {cout << inter << "\t" << templ_mem ; cout <<"\t" << stor_val << endl; }};
template
private: int internal; T temp_var;
public: temp(int i, T j): internal(i),temp_var(j) {cout <<"***Inside C++ temp Constructor" < ~temp() {cout <<"***Inside C++ temp destructor" < void print(void) {cout << internal << "\t" << temp_var << endl;}}; Figure 5. C++ code called from Fortran (cplus.h) Compiling this program, linking it with the bgxlC command, and running it produces the following output: Inside Fortran calling first C function %Inside C function before creating C++ Object ***Inside C++ constructor 10 30 35 %Inside C function after creating C++ Object ***Inside C++ Destructor Inside Fortran calling second C function %Inside C function cfun1 before creating C++ Object ***Inside C++ temp Constructor 40 50.54 %Inside C function after creating C++ temp object ***Inside C++ temp destructor Exiting the Fortran program Making calls to C functions work When you pass an argument to a subprogram call, the usual Fortran convention is to pass the address of the argument. Many C functions expect arguments to be passed as values, however, not as addresses. For these arguments, specify them as %VAL(argument) in the call to C, or make use of the standards-compliant VALUE attribute. For example: 256 XL Fortran: Optimization and Programming Guide MEMBLK = MALLOC(1024) ! Wrong, passes the address of the constant MEMBLK = MALLOC(N) ! Wrong, passes the address of the variable MEMBLK = MALLOC(%VAL(1024)) ! Right, passes the value 1024 MEMBLK = MALLOC(%VAL(N)) ! Right, passes the value of the variable See “Passing arguments by reference or by value” on page 261 and %VAL and %REF in the XL Fortran Language Reference for more details. Passing data from one language to another You need to account for corresponding data types in Fortran and C when passing data from one language to another. The Corresponding data types in Fortran and C table shows the data types available in the XL Fortran and C languages. Further topics detail how Fortran arguments can be passed by reference to C programs. To use the Fortran 2003 Standard interoperability features, see the BIND attribute and ISO_C_BINDING module in the XL Fortran Language Reference. Passing arguments between languages When calling Fortran procedures, the C routines must pass arguments as pointers to the types listed in the following table. Table 23. Corresponding data types in Fortran and C XL Fortran Data Types XL C/C++ Data Types INTEGER(1), BYTE signed char INTEGER(2) signed short INTEGER(4) signed int INTEGER(8) signed long long REAL, REAL(4) float REAL(8), DOUBLE PRECISION double REAL(16) long double (see note 1) COMPLEX, COMPLEX(4) float _Complex COMPLEX(8), DOUBLE COMPLEX double _Complex COMPLEX(16) long double _Complex (see note 1) LOGICAL(1) unsigned char LOGICAL(2) unsigned short LOGICAL(4) unsigned int LOGICAL(8) unsigned long long CHARACTER char CHARACTER(n) char[n] Integer POINTER void * Array array Sequence-derived type structure (with C/C++ -qalign=packed option) Note: 1. Requires C/C++ compiler -qlongdbl option. Chapter 8. Interlanguage calls 257 Notes: 1. In interlanguage communication, it is often necessary to use the %VAL built-in function, or the standards-compliant VALUE attribute, and the %REF built-in function that are defined in “Passing arguments by reference or by value” on page 261. 2. C programs automatically convert float values to double and short integer values to integer when calling an unprototyped C function. Because XL Fortran does not perform a conversion on REAL(4) quantities passed by value, you should not pass REAL(4) and INTEGER(2) by value to a C function that you have not declared with an explicit interface. 3. The Fortran-derived type and the C structure must match in the number, data type, and length of subobjects to be compatible data types. One or more sample programs under the directory /opt/ibmcmp/xlf/bg/14.1/ samples illustrate how to call from Fortran to C. To use the Fortran 2003 Standard interoperability features provided by XL Fortran, see the Language interoperability features section in the XL Fortran Language Reference. Passing global variables between languages To access a C data structure from within a Fortran program or to access a common block from withinaCprogram, follow these steps: 1. Create a named common block that provides a one-to-one mapping of the C structure members. If you have an unnamed common block, change it to a named one. Name the common block with the name of the C structure. 2. Declare the C structure as a global variable by putting its declaration outside any function or inside a function with the extern qualifier. 3. Compile the C source file to get packed structures. program cstruct struct mystuff { real(8) a,d double a; integer b,c int b,c; . double d; .}; common /mystuff/ a,b,c,d . main() { . end } If you do not have a specific need for a named common block, you can create a sequence-derived type with the same one-to-one mapping as a C structure and pass it as an argument to a C function. You must compile the C source file to get packed structures or put #pragmas into the struct. Common blocks that are declared THREADLOCAL are thread-specific data areas that are dynamically allocated by compiler-generated code. A static block is still reserved for a THREADLOCAL common block, but the compiler and the compiler's runtime environment use it for control information. If you need to share THREADLOCAL common blocks between Fortran and C procedures, your C source must be aware of the implementation of the THREADLOCAL common block. For more information, see the Directives section in the XL Fortran Language Reference, and Chapter 12, “Sample Fortran programs,” on page 305. Common blocks that are declared THREADPRIVATE can be accessed using a C global variable that is declared as THREADPRIVATE. 258 XL Fortran: Optimization and Programming Guide Passing character types between languages One difficult aspect of interlanguage calls is passing character strings between languages. The difficulty is due to the following underlying differences in the way that different languages represent such entities: v The only character type in Fortran is CHARACTER, which is stored as a set of contiguous bytes, one character per byte. The length is not stored as part of the entity. Instead, it is passed by value as an extra argument at the end of the declared argument list when the entity is passed as an argument. For the 64-bit complication mode, the size of the argument is 8 bytes. v Character strings in C are stored as arrays of the type char. A null character indicates the end of the string. Note: To have the compiler automatically add the null character to certain character arguments, you can use the -qnullterm option (described in the XL Fortran Compiler Reference. If you are writing both parts of the mixed-language program, you can make the C routines deal with the extra Fortran length argument, or you can suppress this extra argument by passing the string using the %REF function. If you use %REF, which you typically would for pre-existing C routines, you need to indicate where the string ends by concatenating a null character to the end of each character string that is passed toaCroutine: ! Initialize a character string to pass to C. character*6 message1 /’Hello\0’/ ! Initialize a character string as usual, and append the null later. character*5 message2 /’world’/ ! Pass both strings to a C function that takes 2 (char *) arguments. call cfunc(%ref(message1), %ref(message2 // ’\0’)) end For compatibility with C language usage, you can encode the following escape sequences in XL Fortran character strings: Table 24. Escape sequences for character strings Escape Meaning \b Backspace \f Form feed \n New-line \r Carriage return \t Tab \0 Null \' Apostrophe (does not terminate a string) \" Double quotation mark (does not terminate a string) \ \ Backslash \xx, where x is any other character (the backslash is ignored) If you do not want the backslash interpreted as an escape character within strings, you can compile with the -qnoescape option. Chapter 8. Interlanguage calls 259 Passing arrays between languages Fortran stores array elements in ascending storage units in column-major order. C stores array elements in row-major order. Fortran array indexes start at 1, while C array indexes start at 0. The following example shows how a two-dimensional array that is declared by A(3,2) is stored in Fortran and C. Table 25. Corresponding array layouts for Fortran and C. The Fortran array reference A(X,Y,Z) can be expressed in C as a[Z-1][Y-1][X-1]. Keep in mind that although C passes individual scalar array elements by value, it passes arrays by reference. Fortran Element Name C Element Name Lowest storage unit A(1,1) A[0][0] A(2,1) A[0][1] A(3,1) A[1][0] A(1,2) A[1][1] A(2,2) A[2][0] Highest storage unit A(3,2) A[2][1] To pass all or part of a Fortran array to another language, you can use Fortran 90/Fortran 95 array notation: REAL, DIMENSION(4,8) :: A, B(10) ! Pass an entire4x8array. CALL CFUNC( A ) ! Pass only the upper-left quadrant of the array. CALL CFUNC( A(1:2,1:4) ) ! Pass an array consisting of every third element of A. CALL CFUNC( A(1:4:3,1:8) ) ! Pass a 1-dimensional array consisting of elements 1, 2, and 4 of B. CALL CFUNC( B( (/1,2,4/) ) ) Where necessary, the Fortran program constructs a temporary array and copies all the elements into contiguous storage. In all cases, the C routine needs to account for the column-major layout of the array. Any array section or noncontiguous array is passed as the address of a contiguous temporary unless an explicit interface exists where the corresponding dummy argument is declared as an assumed-shape array or a pointer. To avoid the creation of array descriptors (which are not supported for interlanguage calls) when calling non-Fortran procedures with array arguments, either do not give the non-Fortran procedures any explicit interface, or do not declare the corresponding dummy arguments as assumed-shape or pointers in the interface: ! This explicit interface must be changed before the C function ! can be called. INTERFACE FUNCTION CFUNC (ARRAY, PTR1, PTR2) INTEGER, DIMENSION (:) :: ARRAY ! Change this : to *. INTEGER, POINTER, DIMENSION (:) :: PTR1 ! Change this : to * ! and remove the POINTER ! attribute. REAL, POINTER :: PTR2 ! Remove this POINTER ! attribute or change to TARGET. END FUNCTION END INTERFACE 260 XL Fortran: Optimization and Programming Guide Passing pointers between languages Integer POINTERs always represent the address of the pointee object and must be passed by value: CALL CFUNC(%VAL(INTPTR)) Fortran 90 POINTERs can also be passed back and forth between languages but only if there is no explicit interface for the called procedure or if the argument in the explicit interface does not have a POINTER attribute or assumed-shape declarator. You can remove any POINTER attribute or change it to TARGET and can change any deferred-shape array declarator to be explicit-shape or assumed-size. Because of XL Fortran's call-by-reference conventions, you must pass even scalar values from another language as the address of the value, rather than the value itself. For example, a C function passing an integer value x to Fortran must pass &x. Also, a C function passing a pointer value p to Fortran so that Fortran can use it as an integer POINTER must declare it as void **p. A C array is an exception: you can pass it to Fortran without the & operator. Passing arguments by reference or by value To call subprograms written in languages other than Fortran (for example, user-written C programs, or operating system routines), the actual arguments may need to be passed by a method different from the default method used by Fortran. C routines, including those in system libraries, such as libc.so, require you to pass arguments by value instead of by reference. (Although C passes individual scalar array elements by value, it passes arrays by reference.) You can change the default passing method by using the %VAL built-in function or VALUE attribute and the %REF built-in function in the argument list of a CALL statement or function reference. You cannot use them in the argument lists of Fortran procedure references or with alternate return specifiers. %REF Passes an argument by reference (that is, the called subprogram receives the address of the argument). It is the same as the default calling method for Fortran except that it also suppresses the extra length argument for character strings. %VAL Passes an argument by value (that is, the called subprogram receives an argument that has the same value as the actual argument, but any change to this argument does not affect the actual argument). You can use this built-in function with actual arguments that are CHARACTER(1), BYTE, logical, integer, real, or complex expressions or that are sequence-derived type. Objects of derived type cannot contain pointers, arrays, or character structure components whose lengths are greater than one byte. You cannot use %VAL with actual arguments that are array entities, procedure names, or character expressions of length greater than one byte. %VAL causes XL Fortran to pass the actual argument as intermediate values. 64-bit intermediate values If the actual argument is one of the following: Chapter 8. Interlanguage calls 261 v An integer or a logical that is shorter than 64 bits, it is sign-extended to a 64-bit value. v Of type real or complex, it is passed as multiple 64-bit intermediate values. v Of sequence-derived type, it is passed as multiple 64-bit intermediate values. Byte-named constants and variables are passed as if they were INTEGER(1). If the actual argument is a CHARACTER(1), the compiler pads it on the left with zeros to a 64-bit value, regardless of whether you specified the -qctyplss compiler option. If you specified the -qautodbl compiler option, any padded storage space is not passed except for objects of derived type. VALUE attribute Specifies an argument association between a dummy and an actual argument that allows you to pass the dummy argument with the value of the actual argument. Changes to the value or definition status of the dummy argument do not affect the actual argument. You must specify the VALUE attribute for dummy arguments only. You must not use the %VAL or %REF built-in functions to reference a dummy argument with the VALUE attribute, or the associated actual argument. A referenced procedure that has a dummy argument with the VALUE attribute must have an explicit interface. You must not specify the VALUE attribute with the following: v Arrays v Derived types with ALLOCATABLE components v Dummy procedures EXTERNAL FUNC COMPLEX XVAR IVARB=6 CALL RIGHT2(%REF(FUNC)) ! procedure name passed by reference CALL RIGHT3(%VAL(XVAR)) ! complex argument passed by value CALL TPROG(%VAL(IVARB)) ! integer argument passed by value END Explicit interface for %VAL and %REF You can specify an explicit interface for non-Fortran procedures to avoid coding calls to %VAL and %REF in each argument list, as follows: INTERFACE FUNCTION C_FUNC(%VAL(A),%VAL(B)) ! Now you can code "c_func(a,b)" INTEGER A,B ! instead of END FUNCTION C_FUNC ! "c_func(%val(a),%val(b))". END INTERFACE Example with VALUE attribute Program validexm1 integer ::x=10,y=20 print *, ’before calling: ’, x, y call intersub(x, y) print *, ’after calling: ’, x, y contains subroutine intersub(x,y) 262 XL Fortran: Optimization and Programming Guide integer, value :: x integer y x=x+y y = x*y print *, ’in subroutine after changing: ’, x, y end subroutine end program validexm1 Expected output: before calling: 10 20 in subroutine after changing: 30 600 after calling: 10 600 Passing COMPLEX values to/from gcc Passing COMPLEX values between Fortran and GCC depends on what is specified for the -qfloat=[no]complexgcc suboption. If -qfloat=complexgcc is specified, the compiler uses Blue Gene/Q conventions when passing or returning complex numbers. -qfloat=nocomplexgcc is the default. For -qfloat=complexgcc in 64-bit mode, COMPLEX*8 values are passed in 1 GPR, and COMPLEX*16 in 2 GPRs. For -qfloat=nocomplexgcc, COMPLEX*8 and COMPLEX*16 values are passed in 2 floating-point registers (FPRs). COMPLEX*32 values are always passed in 4 FPRs for both -qfloat=complexgcc and -qfloat=nocomplexgcc (since gcc does not support COMPLEX*32). For -qfloat=complexgcc in 64-bit mode, COMPLEX*8 values are returned in GPR3, and COMPLEX*16 in GPR 3-GPR4. For -qfloat=nocomplexgcc, COMPLEX*8 and COMPLEX*16 values are returned in FPR1-FPR2. For both -qfloat=complexgcc and -qfloat=nocomplexgcc, COMPLEX*32 is always returned in FPR1-FPR4. Returning values from Fortran functions XL Fortran does not support calling certain types of Fortran functions from non-Fortran procedures. If a Fortran function returns a pointer, array, or character of nonconstant length, do not call it from outside Fortran. You can call such a function indirectly: SUBROUTINE MAT2(A,B,C) ! You can call this subroutine from C, and the ! result is stored in C. INTEGER, DIMENSION(10,10) :: A,B,C C = ARRAY_FUNC(A,B) ! But you could not call ARRAY_FUNC directly. END Arguments with the OPTIONAL attribute When you pass an optional argument by reference, the address in the argument list is zero if the argument is not present. When you pass an optional argument by value, the value is zero if the argument is not present. The compiler uses an extra register argument to differentiate that value from a regular zero value. If the register has the value 1, the optional argument is present; if it has the value 0, the optional argument is not present. Chapter 8. Interlanguage calls 263 Related information: “Order of arguments in argument list” on page 269 Assembler-level subroutine linkage conventions The subroutine linkage convention specifies the machine state at subroutine entry and exit, allowing routines that are compiled separately in the same or different languages to be linked. The information on subroutine linkage and system calls in the System V Application Binary Interface: PowerPC Processor Supplement and 64–bit PowerPC ELF Application Binary Interface Supplement are the base references on this topic. You should consult these for full details. This section summarizes the information needed to write mixed-language Fortran and assembler programs or to debug at the assembler level, where you need to be concerned with these kinds of low-level details. The system linkage convention passes arguments in registers, taking full advantage of the large number of floating-point registers (FPRs), general-purpose registers (GPRs), vector registers (VPRs) and minimizing the saving and restoring of registers on subroutine entry and exit. The linkage convention allows for argument passing and return values to be in FPRs, GPRs, or both. The following table lists floating-point registers and their functions. The floating-point registers are double precision (64 bits). Table 26. Floating-point register usage across calls Register Preserved Across Calls Use 0no 1 no FP parameter 1, function return 1. 2 no FP parameter 2, function return 2. 3 no FP parameter 3, function return complex *32. 4 no FP parameter 4, function return complex *32...... 8 no FP parameter 8 9-13 no 14-31 yes local variables The following table lists general-purpose registers and their functions. Table 27. General-purpose register usage across calls Register Preserved Across Calls Use 0no 1 yes Stack pointer. 2 yes System-reserved. 3 no 1st word of arg list; return value 1. 4 no 2nd word of arg list; return value 2. 264 XL Fortran: Optimization and Programming Guide Table 27. General-purpose register usage across calls (continued) Register Preserved Across Calls Use 5 no 3rd word of arg list ...... 10 no 8th word of arg list. 11-12 no 13 yes SDA pointer. 14-30 no Local variables. 31 yes Local variables or “environment pointers”. If a register is not designated as preserved, its contents may be changed during the call, and the caller is responsible for saving any registers whose values are needed later. Conversely, if a register is supposed to be preserved, the callee is responsible for preserving its contents across the call, and the caller does not need any special action. The following table lists special-purpose register conventions. Table 28. Special-purpose register usage across calls Register Preserved Across Calls Condition register Bits 0-7 (CR0,CR1) no Bits 8-22 (CR2,CR3,CR4) yes Bits 23-31 (CR5,CR6,CR7) no Link register no Count register no XER register no FPSCR register no The stack The stack is a portion of storage that is used to hold local storage, register save areas, parameter lists, and call-chain data. The stack grows from higher addresses to lower addresses. A stack pointer register (register 1) is used to mark the current “top” of the stack. A stack frame is the portion of the stack that is used by a single procedure. The input parameters are considered part of the current stack frame. In a sense, each output argument belongs to both the caller's and the callee's stack frames. In either case, the stack frame size is best defined as the difference between the caller's stack pointer and the callee's. The following diagram shows the storage map of typical stack frames for 64-bit environments. In this diagram, the current routine has acquired a stack frame that allows it to call other functions. If the routine does not make any calls and there are no local variables or temporaries, and it does not need to save any non-volatile registers, the function need not allocate a stack frame. It can still use the register save area at the top of the caller's stack frame, if needed. Chapter 8. Interlanguage calls 265 The stack frame is double-word aligned. LOW Stack grows at ADDRESSES this end Callee's stack 0 Back chain pointer 8 Saved CR 16 Saved LR 24-32 Reserved LINK AREA 40 Saved TOC (callee) OUTPUT ARGUMENT AREA P1 Space for P1-P8 (Used by callee is always reserved to construct Pn argument list) Callee's stack LOCAL STACK AREA area (Possible word wasted for alignment.) -8*nfprs-8*ngprs Save area for Rfirst = R13 for full save caller's GPR max save 19 doublewords R31 -8*nfprs Save area for Ffirst = F14 for a caller's FPR max full save 18 dblwds F31 Caller's stack 0 Back chain pointer 8 Saved CR 16 Saved LR 24-32 Reserved LINK AREA 40 Saved TOC (caller) Space for P1-P8 48 P1 INPUT PARAMETER AREA is always reserved (Callee's input Pn parameters found here. Is also caller's arg area.) Caller's HIGH stack ADDRESSES area Figure 6. Runtime Stack for 64-bit Environment The Linkage Area and Minimum Stack Frame In a 64-bit environment, the linkage area consists of six doublewords at offset zero from the caller's stack pointer on entry to a procedure. The first doubleword contains the caller's back chain (stack pointer). The second doubleword is the location where the callee saves the Condition Register (CR) if it is needed. The third doubleword is the location where the callee's prolog code saves the Link Register if it is needed. The fourth doubleword is reserved for C SETJMP and LONGJMP processing, and the fifth doubleword is reserved for future use. The last doubleword (doubleword 6) is reserved for use by the global linkage routines that are used when calling routines in other object modules (for example, in shared libraries). 266 XL Fortran: Optimization and Programming Guide The input parameter area In a 64-bit environment, the input parameter area is a contiguous piece of storage reserved by the calling program to represent the register image of the input parameters of the callee. The input parameter area is double-word aligned and is located on the stack directly following the caller's link area. This area is at least 8 doublewords in size. If more than 8 doublewords of parameters are expected, they are stored as register images that start at positive offset 112 from the incoming stack pointer. The first 8 doublewords only appear in registers at the call point, never in the stack. Remaining words are always in the stack, and they can also be in registers. The register save area In a 64-bit environment, the register save area is double-word aligned. It provides the space that is needed to save all nonvolatile FPRs and GPRs used by the callee program. The FPRs are saved next to the link area. The GPRs are saved below the FPRs (in lower addresses). The called function may save the registers here even if it does not need to allocate a new stack frame. The system-defined stack floor includes the maximum possible save area: 18*8 for FPRs + 19*4 for GPRs 64-bit platforms: 18*8 for FPRs + 19*8 for GPRs A callee needs only to save the nonvolatile registers that it actually uses. The local stack area The local stack area is the space that is allocated by the callee procedure for local variables and temporaries. The output parameter area In a 64-bit environment, the output parameter area (P1...Pn) must be large enough to hold the largest parameter list of all procedures that the procedure that owns this stack frame calls. This area is at least 8 doublewords long, regardless of the length or existence of any argument list. If more than 8 doublewords are being passed, an extension list is constructed, which begins at offset 112 from the current stack pointer. The first 8 doublewords only appear in registers at the call point, never in the stack. Remaining doublewords are always in the stack, and they can also be in registers. Linkage convention for argument passing The system linkage convention takes advantage of the large number of registers available. The linkage convention passes arguments in both GPRs and FPRs. Two fixed lists, R3-R10 and FP1-FP13, specify the GPRs and FPRs available for argument passing. When there are more argument words than available argument GPRs and FPRs, the remaining words are passed in storage on the stack. The values in storage are the same as if they were in registers. Chapter 8. Interlanguage calls 267 In a 64-bit environment, the size of the parameter area is sufficient to contain all the arguments passed on any call statement from a procedure that is associated with the stack frame. Although not all the arguments for a particular call actually appear in storage, it is convenient to consider them as forming a list in this area, each one occupying one or more words. For call by reference (as is the default for Fortran), the address of the argument is passed in a register. The following information refers to call by value, as in C or as in Fortran when %VAL is used. For purposes of their appearance in the list, arguments are classified as floating-point values or non-floating-point values: v All non-floating-point values require one doubleword that is doubleword aligned. v Each single-precision (REAL(4)) value and each double-precision (REAL(8)) value occupies one doubleword in the list. Each extended-precision (REAL(16)) value occupies two successive doublewords in the list. v A COMPLEX value occupies twice as many doublewords as a REAL value with the same kind type parameter. v In Fortran and C, structure values appear in successive words as they would anywhere in storage, satisfying all appropriate alignment requirements. Structures are aligned to a doubleword and occupy (sizeof(struct X)+7)/8 doublewords, with any padding at the end. A structure that is smaller than a doubleword is left-justified within its doubleword or register. Larger structures can occupy multiple registers and may be passed partly in storage and partly in registers. v Other aggregate values are passed “val-by-ref”. That is, the compiler actually passes their address and arranges for a copy to be made in the invoked program. v A procedure or function pointer is passed as a pointer to the routine's function descriptor; its first word contains its entry point address. (See “Pointers to functions” on page 270 for more information.) Argument passing rules (by value) In a 64-bit environment, from the following illustration, we state these rules: v If the called procedure treats the parameter list as a contiguous piece of storage (for example, if the address of a parameter is taken in C), the parameter registers are stored in the space reserved for them in the stack. v A register image is stored on the stack. v The argument area (P1...Pn) must be large enough to hold the largest parameter list. Here is an example of a call to a function : f(%val(l1), %val(l2), %val(l3), %val(l4), %val(l5), %val(l6), %val(l7), %val(d1), %val(f1), %val(c1), %val(d2), %val(s1), %val(cx2)) where: l denotes integer(4) (fullword integer) d denotes real(8) (double precision) f denotes real(4) (real) s denotes integer(2) (halfword integer) c denotes character (one character) cx denotes complex(8) (double complex) 268 XL Fortran: Optimization and Programming Guide Storage Mapping of Parm Area On the Stack in Will Be Passed In: 64-Bit Environment R3 0 L1 R4 8 L2 R5 16 L3 R6 24 L4 R7 32 L5 R8 40 L6 R9 48 L7 FP1 56 D1 R10 unused FP2 64 F1 stack 72 C1 right justified (if language semantics specify) FP3 80 D2 stack 88 S1 right justified (if language semantics specify) FP4 96 CX2(real) FP5 104 CX2(imaginary) Figure 7. Storage mapping of parm area on the stack in 64-bit environment Order of arguments in argument list The argument list is constructed in the following order. Items in the same bullet appear in the same order as in the procedure declaration, whether or not argument keywords are used in the call. v All addresses or values (or both) of actual arguments 1 v “Present” indicators for optional arguments v Length arguments for strings 1 Linkage convention for function calls Function calls to a routine make use of its function descriptor and entry point symbols. In 64-bit mode, a routine has two symbols associated with it: a function descriptor (name) and an entry point (.name). When a call is made to a routine, the program branches to the entry point directly. Excluding the loading of parameters (if any) in the proper registers, compilers expand calls to functions to the following two-instruction sequence: BL .foo # Branch to foo ORI R0,R0,0x0000 # Special NOP The linker does one of two things when it encounters a BL instruction: 1. There may be other items in this list during Fortran-Fortran calls. However, they will not be visible to non-Fortran procedures that follow the calling rules in this section. Chapter 8. Interlanguage calls 269 1. If foo is imported (not in the same object module), the linker changes the BL to .foo to a BL to .glink (global linkage routine) of foo and inserts the .glink into the object module. Also, if a NOP instruction (ORI R0,R0,0x0000) immediately follows the BL instruction, the linker replaces the NOP instruction with the LOAD instruction L R2, 20(R1). 2. If foo is bound in the same object module as its caller and a LOAD instruction L R2,40(R1) for 64-bit, or ORI R0,R0,0 immediately follows the BL instruction, the linker replaces the LOAD instruction with a NOP (ORI R0,R0,0). Note: For any export, the linker inserts the procedure's descriptor into the object module. Pointers to functions In 64–bit mode, a function pointer is a data type whose values range over procedure names. Variables of this type appear in several programming languages, such as C and Fortran. In Fortran, a dummy argument that appears in an EXTERNAL statement is a function pointer. Fortran provides support for the use of function pointers in contexts such as the target of a call statement or an actual argument of such a statement. A function pointer is a fullword quantity that is the address of a function descriptor. The function descriptor is a 3-word object. The first word contains the address of the entry point of the procedure. The second has the address of the TOC of the object module in which the procedure is bound. The third is the environment pointer for some non-Fortran languages. There is only one function descriptor per entry point. It is bound into the same object module as the function it identifies if the function is external. The descriptor has an external name, which is the same as the function name but with a different storage class that uniquely identifies it. This descriptor name is used in all import or export operations. In 64–bit mode, function pointers are 8 bytes long and contain a 64-bit address. For pointers to local functions, the address contained is the address of the function in the text section. For imported functions, the address is that of the function's stub. Every unique, imported function will have a stub in the object. The function stub is in the non-lazy symbol pointer section. Function values Functions return their values according to type: v INTEGER and LOGICAL of kind 1, 2, and 4 are returned (right justified) in R3. v REAL(4) or REAL(8) are returned in FP1. REAL(16) are returned in FP1 and FP2. v COMPLEX(4) or COMPLEX(8) are returned in FP1 and FP2. COMPLEX(16) are returned in FP1-FP4. v v Vector results are returned in VPR2 v Character strings are returned in a buffer allocated by the caller. The address and the length of this buffer are passed in R3 and R4 as hidden parameters. The first explicit parameter word is in R5, and all subsequent parameters are moved to the next word. v Structures are returned in a buffer that is allocated by the caller. The address is passed in R3; there is no length. The first explicit parameter is in R4. 270 XL Fortran: Optimization and Programming Guide The stack floor In 64–bit mode, the stack floor is a system-defined address below which the stack cannot grow. All programs in the system must avoid accessing locations in the stack segment that are below the stack floor. All programs must maintain other system invariants that are related to the stack: v No data is saved or accessed from an address lower than the stack floor. v The stack pointer is always valid. When the stack frame size is more than 32 767 bytes, you must take care to ensure that its value is changed in a single instruction. This step ensures that there is no timing window where a signal handler would either overlay the stack data or erroneously appear to overflow the stack segment. Stack overflow The linkage convention requires no explicit inline check for overflow. The operating system uses a storage protection mechanism to detect stores past the end of the stack segment. Prolog and epilog You need to consider a number of steps when entering a procedure and when exiting a procedure. On entry to a procedure, you might have to do some or all of the following steps: 1. Save the link register. 2. If you use any of the CR bits 8-2319 (CR2, CR3, CR4, CR5), save the CR. 3. Save any nonvolatile FPRs that are used by this procedure in the FPR save area. 4. Save all nonvolatile VPRs that are used by this procedure in the callers VPR save area. 5. Save the VRSAVE register 6. Save all nonvolatile GPRs that are used by this procedure in the GPR save area. 7. Store back chain and decrement stack pointer by the size of the stack frame. Note that if a stack overflow occurs, it will be known immediately when the store of the back chain is done. On exit from a procedure, you might have to perform some or all of the following steps: 1. Restore all GPRs saved. 2. Restore all VPRs saved 3. Restore the VRSAVE register 4. Restore stack pointer to the value it had on entry. 5. Restore link register if necessary. 6. Restore bits 8-2319 of the CR if necessary. 7. If you saved any FPRs, restore them. 8. Return to caller. Chapter 8. Interlanguage calls 271 Traceback In 64–bit mode, the compiler supports the traceback mechanism, which symbolic debuggers need to unravel the call or return stack. Each object module has a traceback table in the text segment at the end of its code. This table contains information about the object module, including the type of object module, as well as stack frame and register information. Note: You can make the traceback table smaller or remove it entirely with the -qtbtable option. 272 XL Fortran: Optimization and Programming Guide Chapter 9. Implementation details of XL Fortran Input/Output (I/O) This topic describes XL Fortran support (through extensions and platform-specific details) for the Blue Gene/Q file system. See “Mixed-language input and output” on page 254 for further considerations related to input and output operations. Implementation details of file formats The manner in which XL Fortran implements files is based on their file format. Sequential-access unformatted files: An integer that contains the length of the record precedes and follows each record. For 64-bit applications, the length of the integer is 4 bytes if you set the uwidth runtime option to 32 (the default), and 8 bytes if you set the uwidth runtime option to 64. Sequential-access formatted files: XL Fortran programs break these files into records while reading, by using each newline character (X'0A') as a record separator. On output, the input/output system writes a newline character at the end of each record. Programs can also write newline characters for themselves. This practice is not recommended because the effect is that the single record that appears to be written is treated as more than one record when being read or backspaced over. Direct access files: XL Fortran simulates direct-access files with operating system files whose length is a multiple of the record length of the XL Fortran file. You must specify, in an OPEN statement, the record length (RECL)ofthe direct-access file. XL Fortran uses this record length to distinguish records from each other. For example, the third record of a direct-access file of record length 100 bytes would start at the 201st byte of the single record of a Blue Gene/Q file and end at the 300th byte. If the length of the record of a direct-access file is greater than the total amount of data you want to write to the record, XL Fortran pads the record on the right with blanks (X'20'). Stream-access unformatted files: Unformatted stream files are viewed as a collection of file storage units. In XL Fortran, a file storage unit is one byte. A file connected for unformatted stream access has the following properties: v The first file storage unit has position 1. Each subsequent file storage unit has a position that is one greater than that of the preceding one. v For a file that can be positioned, file storage units need not be read or written in the order of their position. Any file storage unit may be read from the file while it is connected to a unit, provided that the file © Copyright IBM Corp. 1990, 2012 273 storage unit has been written since the file was created, and if a READ statement for the connection is permitted. Stream-access formatted files: A record file connected for formatted stream access has the following properties: v Some file storage units may represent record markers. The record marker is the newline character (X'0A'). v The file will have a record structure in addition to the stream structure. v The record structure is inferred from the record markers that are stored in the file. v Records can have any length up to the internal limit allowed by XL Fortran (See XL Fortran Internal limits in the XL Fortran Compiler Reference.) v There may or may not be a record marker at the end of the file. If there is no record marker at the end of the file, the final record is incomplete, but not empty. A file connected for formatted stream access has the following properties: v The first file storage unit has position 1. Each subsequent file storage unit has a position that is greater than that of the preceding one. Unlike unformatted stream access, the positions of successive file storage units are not always consecutive. v The position of a file connected for formatted stream access can be determined by the POS= specifier in an INQUIRE statement. v For a file that can be positioned, the file position can be set to a value that was previously identified by the POS= specifier in INQUIRE. File names There are a number of considerations to be aware of when working with file names. You can specify file names as either relative (such as file, dir/file,or../file)or absolute (such as /file or /dir/file). The maximum length of a file name (the full path name) is 4095 characters, even if you only specify a relative path name in the I/O statement. The maximum length of a file name with no path is 255 characters. You must specify a valid file name in such places as the following: v The FILE= specifier of the OPEN and INQUIRE statements v INCLUDE lines Note: To specify a file whose location depends on an environment variable, you can use the GET_ENVIRONMENT_VARIABLE intrinsic procedure to retrieve the value of the environment variable: character(100) home, name call get_environment_variable(’HOME’, value=home) ! Now home = $HOME + blank padding. ! Construct the complete path name and open the file. name=trim(home) // ’/remainder/of/path’ open (unit=10, file=name) ... end 274 XL Fortran: Optimization and Programming Guide Preconnected and Implicitly Connected Files Whether files are preconnected or implicitly connected files is dependent on their units and specific statements. Units 0, 5, and 6 are preconnected to standard error, standard input, and standard output, respectively, before the program runs. All other units can be implicitly connected when an ENDFILE, PRINT, READ, REWIND,orWRITE statement is performed on a unit that has not been opened. Unit n is implicitly connected to a file that is named fort.n. These files need not exist, and XL Fortran does not create them unless you use the corresponding units implicitly. Note: Because unit 0 is preconnected for standard error, you cannot use it for the following statements: CLOSE, ENDFILE, BACKSPACE, REWIND, and direct or stream input/output. You can use it in an OPEN statement only to change the values of the BLANK=, DELIM=, DECIMAL=or PAD= specifiers. You can also implicitly connect units 5 and 6 (and *) by using I/O statements that follow a CLOSE of these units: WRITE (6,10) "This message goes to stdout." CLOSE (6) WRITE (6,10) "This message goes in the file fort.6." PRINT *, "Output to * now also goes in fort.6." 10 FORMAT (A) END The FORM= specifier of implicitly connected files has the value FORMATTED before any READ, WRITE,orPRINT statement is performed on the unit. The first such statement on such a file determines the FORM= specifier from that point on: FORMATTED if the formatting of the statement is format-directed, list-directed, or namelist; and UNFORMATTED if the statement is unformatted. Preconnected files also have FORM='FORMATTED', STATUS='OLD', and ACTION='READWRITE' as default specifier values. The other properties of a preconnected or implicitly connected file are the default specifier values for the OPEN statement. These files always use sequential access. If you want XL Fortran to use your own file instead of the fort.n file, you can either specify your file for that unit through an OPEN statement or create a symbolic link before running the application. In the following example, there is a symbolic link between myfile and fort.10: ln -s myfile fort.10 When you run an application that uses the implicitly connected file fort.10 for input/output, XL Fortran uses the file myfile instead. The file fort.10 exists, but only as a symbolic link. The following command will remove the symbolic link, but will not affect the existence of myfile: rm fort.10 Chapter 9. Implementation details of XL Fortran Input/Output (I/O) 275 File positioning The position of a file pointer when a file is opened with no POSITION= specifier is summarized in the following table. Table 29. Position of the file pointer when a file is opened with no POSITION= specifier -qposition suboptions Implicit OPEN Explicit OPEN STATUS = STATUS = 'OLD' STATUS = 'NEW' 'UNKNOWN' File File File File File File File File exists does exists does exists does exists does not not not not exist exist exist exist option not specified Start Start Error Start Start Error Start Start appendold Start Start Error Start End Error Start Start appendunknown Start Start Error Start Start Error End Start appendold and appendunknown Start Start Error Start End Error End Start I/O redirection You can use the redirection operator on the command line to redirect input to and output from your XL Fortran program. How you specify and use this operator depends on which shell you are running. Hereisabash example: $ cat redirect.f write (6,*) ’This goes to standard output’ write (0,*) ’This goes to standard error’ read (5,*) i print *,i end $ bgxlf95 redirect.f ** _main === End of Compilation 1 === 1501-510 Compilation successful for file redirect.f. $ # No redirection. Input comes from the terminal. Output goes to $ # the screen. $ a.out This goes to standard output This goes to standard error 4 4 $ # Create an input file. $ echo >stdin 2 $ # Redirect each standard I/O stream. $ a.out >stdout 2>stderr Refer to your man pages for more information on redirection. 276 XL Fortran: Optimization and Programming Guide How XL Fortran I/O interacts with pipes, special files, and links You can access regular operating system files and blocked special files by using sequential-access, direct-access, or stream-access methods. You can only access pseudo-devices, pipes, and character special files by using sequential-access methods, or stream-access without using the POS= specifier. When you use symbolic link to link files together, you can use their names interchangeably, as shown in the following example: OPEN (4, FILE="file1") OPEN (4, FILE="link_to_file1", PAD="NO") ! Modify connection Do not specify the POSITION= specifier as REWIND or APPEND for pipes. Do not specify ACTION='READWRITE' for a pipe. Do not use the BACKSPACE statement on files that are pseudo-devices or character special files. Do not use the REWIND statement on files that are pseudo-devices or pipes. Default record lengths The default record lengths for files is dependent on the file format and on the RECL= qualifier. If a pseudo-device, pipe, or character special file is connected for formatted or unformatted sequential access with no RECL= qualifier, or for formatted stream access, the default record length is 32 768 rather than 2 147 483 647, which is the default for sequential-access files connected to random-access devices. (See the default_recl runtime option in the XL Fortran Compiler Reference.) In certain cases, the default maximum record length for formatted files is larger, to accommodate programs that write long records to standard output. If a unit is connected to a terminal for formatted sequential access and there is no explicit RECL= qualifier in the OPEN statement, the program uses a maximum record length of 2 147 483 646 (2**31-2) bytes, rather than the usual default of 32 768 bytes. When the maximum record length is larger, formatted I/O has one restriction: WRITE statements that use the T or TL edit descriptors must not write more than 32 768 bytes. This is because the unit's internal buffer is flushed each 32 768 bytes, and the T or TL edit descriptors will not be able to move back past this boundary. File permissions A file must have the appropriate permissions (read, write, or both) for the corresponding operation being performed on it. When a file is created, the default permissions (if the umask setting is 000) are both read and write for user, group, and other. You can turn off individual permission bits by changing the umask setting before you run the program. Chapter 9. Implementation details of XL Fortran Input/Output (I/O) 277 Selecting error messages and recovery actions There are various ways to control a program's behavior when errors are encountered. By default, an XL Fortran-compiled program continues after encountering many kinds of errors, even if the statements have no ERR= or IOSTAT= specifiers. The program performs some action that might allow it to recover successfully from the bad data or other problem. To control the behavior of a program that encounters errors, set the XLFRTEOPTS environment variable, which is described in Setting runtime options in the XL Fortran Compiler Reference, before running the program: v To make the program stop when it encounters an error instead of performing a recovery action, include err_recovery=no in the XLFRTEOPTS setting. v To make the program stop issuing messages each time it encounters an error, include xrf_messages=no. v To disallow XL Fortran extensions to Fortran 90 at run time, include langlvl=90std. To disallow XL Fortran extensions to Fortran 95 at run time, include langlvl=95std. To disallow XL Fortran extensions to Fortran 2003 behavior at run time, include langlvl=2003std. To disallow XL Fortran extensions to Fortran 2008 behavior at run time, include langlvl=2008std. These settings, in conjunction with the -qlanglvl compiler option, can help you locate extensions when preparing to port a program to another platform. For example: # Switch defaults for some runtime settings. XLFRTEOPTS="err_recovery=no:cnverr=no" export XLFRTEOPTS If you want a program always to work the same way, regardless of environment-variable settings, or want to change the behavior in different parts of the program, you can call the SETRTEOPTS procedure: PROGRAM RTEOPTS USE XLFUTILITY CALL SETRTEOPTS("err_recovery=no") ! Change setting. ... some I/O statements ... CALL SETRTEOPTS("err_recovery=yes") ! Change it back. ... some more I/O statements ... END Because a user can change these settings through the XLFRTEOPTS environment variable, be sure to use SETRTEOPTS to set all the runtime options that might affect the desired operation of the program. Flushing I/O buffers To protect data from being lost if a program ends unexpectedly, you can use the FLUSH statement or the flush_ subroutine to write any buffered data to a file. The FLUSH statement is recommended for better portability and is used in the following example: INTEGER, PARAMETER :: UNIT = 10 DOI=1,1000000 WRITE(UNIT, *) I CALL MIGHT_CRASH ! If the program ends in the middle of the loop, some data 278 XL Fortran: Optimization and Programming Guide ! may be lost. END DO DOI=1,1000000 WRITE(UNIT, *) I FLUSH(UNIT) CALL MIGHT_CRASH ! If the program ends in the middle of the loop, all data written ! up to that point will be safely in the file. END DO END Related information: “Mixed-language input and output” on page 254 See FLUSH in the Compiler Reference Choosing locations and names for Input/Output files If you need to override the default locations and names for input/output files, you can use the following methods without making any changes to the source code. Naming files that are connected with no explicit name To give a specific name to a file that would usually have a name of the form fort.unit, you must set the runtime option unit_vars and then set an environment variable with a name of the form XLFUNIT_unit for each scratch file. The association is between a unit number in the Fortran program and a path name in the file system. For example, suppose that the Fortran program contains the following statements: OPEN (UNIT=1, FORM=’FORMATTED’, ACCESS=’SEQUENTIAL’, RECL=1024) ... OPEN (UNIT=12, FORM=’UNFORMATTED’, ACCESS=’DIRECT’, RECL=131072) ... OPEN (UNIT=123, FORM=’UNFORMATTED’, ACCESS=’SEQUENTIAL’, RECL=997) XLFRTEOPTS="unit_vars=yes" # Allow overriding default names. XLFUNIT_1="/tmp/molecules.dat" # Use this named file. XLFUNIT_12="../data/scratch" # Relative to current directory. XLFUNIT_123="/home/user/data/Users/username/data" # Somewhere besides /tmp. export XLFRTEOPTS XLFUNIT_1 XLFUNIT_12 XLFUNIT_123 Notes: 1. The XLFUNIT_number variable name must be in uppercase, and number must not have any leading zeros. 2. unit_vars=yes might be only part of the value for the XLFRTEOPTS variable, depending on what other runtime options you have set. See Setting runtime options in the XL Fortran Compiler Reference for other options that might be part of the XLFRTEOPTS value. 3. If the unit_vars runtime option is set to no or is undefined or if the applicable XLFUNIT_number variable is not set when the program is run, the program uses a default name (fort.unit) for the file and puts it in the current directory. Naming scratch files To place all scratch files in a particular directory, set the TMPDIR environment variable to the name of the directory. The program then opens the scratch files in this directory. You might need to do this if your /tmp directory is too small to hold the scratch files. Chapter 9. Implementation details of XL Fortran Input/Output (I/O) 279 To give a specific name to a scratch file, you must do the following: 1. Set the runtime option scratch_vars. 2. Set an environment variable with a name of the form XLFSCRATCH_unit for each scratch file. The association is between a unit number in the Fortran program and a path name in the file system. In this case, the TMPDIR variable does not affect the location of the scratch file. For example, suppose that the Fortran program contains the following statements: OPEN (UNIT=1, STATUS=’SCRATCH’, & FORM=’FORMATTED’, ACCESS=’SEQUENTIAL’, RECL=1024) ... OPEN (UNIT=12, STATUS=’SCRATCH’, & FORM=’UNFORMATTED’, ACCESS=’DIRECT’, RECL=131072) ... OPEN (UNIT=123, STATUS=’SCRATCH’, & FORM=’UNFORMATTED’, ACCESS=’SEQUENTIAL’, RECL=997) XLFRTEOPTS="scratch_vars=yes" # Turn on scratch file naming. XLFSCRATCH_1="/tmp/molecules.dat" # Use this named file. XLFSCRATCH_12="../data/scratch" # Relative to current directory. XLFSCRATCH_123="/home/user/data/Users/username/data" # Somewhere besides /tmp. export XLFRTEOPTS XLFSCRATCH_1 XLFSCRATCH_12 XLFSCRATCH_123 Notes: 1. The XLFSCRATCH_number variable name must be in uppercase, and number must not have any leading zeros. 2. scratch_vars=yes might be only part of the value for the XLFRTEOPTS variable, depending on what other runtime options you have set. See Setting runtime options in the XL Fortran Compiler Reference for other options that might be part of the XLFRTEOPTS value. 3. If the scratch_vars runtime option is set to no or is undefined or if the applicable XLFSCRATCH_number variable is not set when the program is run, the program chooses a unique file name for the scratch file and puts it in the directory named by the TMPDIR variable or in the /tmp directory if the TMPDIR variable is not set. XL Fortran thread-safe I/O library The XL Fortran runtime library libxlf90_r.so provides support for parallel execution of Fortran I/O statements. Synchronization of I/O operations During parallel execution, multiple threads might perform I/O operations on the same file at the same time. If they are not synchronized, the results of these I/O operations could be shuffled or merged or both, and the application might produce incorrect results or even terminate. The XL Fortran runtime library synchronizes I/O operations for parallel applications. It performs the synchronization within the I/O library, and it is transparent to application programs. The purpose of the synchronization is to ensure the integrity and correctness of each individual I/O operation. However, the runtime does not have control over the order in which threads execute I/O statements. Therefore, the order of records read in or written out is not predictable under parallel I/O operations. Refer to “Parallel I/O issues” on page 281 for details. 280 XL Fortran: Optimization and Programming Guide External files For external files, the synchronization is performed on a per-unit basis. The XL Fortran runtime ensures that only one thread can access a particular logical unit to prevent several threads from interfering with each other. When a thread is performing an I/O operation on a unit, other threads attempting to perform I/O operations on the same unit must wait until the first thread finishes its operation. Therefore, the execution of I/O statements by multiple threads on the same unit is serialized. However, the runtime environment does not prevent threads from operating on different logical units in parallel. In other words, parallel access to different logical units is not necessarily serialized. Functionality of I/O under synchronization The XL Fortran runtime sets its internal locks to synchronize access to logical units. This should not have any functional impact on the I/O operations performed by a Fortran program. Also, it will not impose any additional restrictions to the operability of Fortran I/O statements except for the use of I/O statements in a signal handler that is invoked asynchronously. Refer to “Use of I/O statements in signal handlers” on page 283 for details. Parallel I/O issues The order in which parallel threads perform I/O operations is not predictable. The XL Fortran runtime does not have control over the ordering. It will allow whichever thread that executes an I/O statement on a particular logical unit and obtains the lock on it first to proceed with the operation. Therefore, only use parallel I/O in cases where at least one of the following is true: v Each thread performs I/O on a predetermined record in direct-access files. v Each thread performs I/O on a different part of a stream-access file. Different I/O statements cannot use the same, or overlapping, areas of a file. v The result of an application does not depend on the order in which records are written out or read in. v Each thread performs I/O on a different file. In these cases, results of the I/O operations are independent of the order in which threads execute. However, you might not get the performance improvements that you expect, since the I/O library serializes parallel access to the same logical unit from multiple threads. Examples of these cases are as follows: v Each thread performs I/O on a pre-determined record in a direct-access file: doi=1,10 write(4, ’(i4)’, rec = i) a(i) enddo v Each thread performs I/O on a different part of a stream-access file. Different I/O statements cannot use the same, or overlapping, areas of a file. doi=1,9 write(4, ’(i4)’, pos=1+5*(i-1))a(i) ! We use 5 above because i4 takes 4 file storage ! units + 1 file storage unit for the record marker. enddo v In the case that each thread operates on a different file, since threads share the status of the logical units connected to the files, the thread still needs to obtain the lock on the logical unit for either retrieving or updating the status of the logical unit. However, the runtime allows threads to perform the data transfer between the logical unit and the I/O list item in parallel. If an application contains a large number of small I/O requests in a parallel region, you might not get the expected performance because of the lock contention. Consider the following example: Chapter 9. Implementation details of XL Fortran Input/Output (I/O) 281 program example use omp_lib integer, parameter :: num_of_threads = 4, max = 5000000 character*10 file_name integer i, file_unit, thread_id integer, dimension(max, 2 * num_of_threads) :: aa call omp_set_num_threads(num_of_threads) !$omp parallel private(file_name, thread_id, file_unit, i) shared(aa) thread_id = omp_get_thread_num() file_name = ’file_’ file_name(6:6) = char(ichar(’0’) + thread_id) file_unit = 10 + thread_id open(file_unit, file = file_name, status = ’old’, action = ’read’) doi=1,max read(file_unit, *) aa(i, thread_id*2+1),aa(i, thread_id*2+2) end do close(file_unit) !$omp end parallel end The XL Fortran runtime synchronizes retrieving and updating the status of the logical units while performing data transfer in parallel. In order to increase performance, it is recommended to increase the size of data transfer per I/O request. The do loop, therefore, should be rewritten as follows: read(file_unit, *) a(:, thread_id*2+1:thread_id*2+2) doi=1,max ! Do something for each element of array ’aa’. end do v The result does not depend on the order in which records are written out or read in: real a(100) doi=1,10 read(4) a(i) enddo call qsort_(a) v Each thread performs I/O on a different logical unit of direct access, sequential access, or stream access: do i = 11, 20 write(i, ’(i4)’) a(i - 10) enddo For multiple threads to write to or read from the same sequential-access file, or to write to or read from the same stream-access file without using the POS= specifier, the order of records written out or read in depends on the order in which the threads execute the I/O statement on them. This order, as stated previously, is not predictable. Therefore, the result of an application could be incorrect if it assumes records are sequentially related and cannot be arbitrarily written out or read in. For example, if the following loop is parallelized, the numbers printed out will no longer be in the sequential order from 1 to 500 as the result of a serial execution: 282 XL Fortran: Optimization and Programming Guide doi=1,500 print *, i enddo Applications that depend on numbers being strictly in the specified order will not work correctly. The XL Fortran runtime option multconn=yes allows connection of the same file to more than one logical unit simultaneously. Since such connections can only be made for reading (ACCESS='READ'), access from multiple threads to logical units that are connected to the same file will produce predictable results. Use of I/O statements in signal handlers There are basically two kinds of signals in the POSIX signal model: synchronously and asynchronously generated signals. Signals caused by the execution of some code of a thread, such as a reference to an unmapped, protected, or bad memory (SIGSEGV or SIGBUS), floating-point exception (SIGFPE), execution of a trap instruction (SIGTRAP), or execution of illegal instructions (SIGILL) are said to be synchronously generated. Signals may also be generated by events outside the process: for example, SIGINT, SIGHUP, SIGQUIT, SIGIO, and so on. Such events are referred to as interrupts. Signals that are generated by interrupts are said to be asynchronously generated. The XL Fortran runtime is asynchronous signal unsafe. This means that an XL Fortran I/O statement cannot be used in a signal handler that is entered because of an asynchronously generated signal. The behavior of the system is undefined when an XL Fortran I/O statement is called from a signal handler that interrupts an I/O statement. However, it is safe to use I/O statements in signal handlers for synchronous signals. Sometimes an application can guarantee that a signal handler is not entered asynchronously. For example, an application might mask signals except when it runs certain known sections of code. In such situations, the signal will not interrupt any I/O statements and other asynchronous signal unsafe functions. Therefore, you can still use Fortran I/O statements in an asynchronous signal handler. A much easier and safer way to handle asynchronous signals is to block signals in all threads and to explicitly wait (using sigwait()) for them in one or more separate threads. The advantage of this approach is that the handler thread can use Fortran I/O statements as well as other asynchronous signal unsafe routines. Asynchronous thread cancellation When a thread enables asynchronous thread cancellability, any cancellation request is acted upon immediately. The XL Fortran runtime environment is not asynchronous thread cancellation safe. The behavior of the system is undefined if a thread is cancelled asynchronously while it is in the XL Fortran runtime environment. Chapter 9. Implementation details of XL Fortran Input/Output (I/O) 283 284 XL Fortran: Optimization and Programming Guide Chapter 10. Implementation details of XL Fortran floating-point processing This topic answers some common questions about floating-point processing. v How can I get predictable, consistent results? v How can I get the fastest or the most accurate results? v How can I detect, and possibly recover from, exception conditions? v Which compiler options can I use for floating-point calculations? The topics describing floating-point precision make frequent reference to the compiler options that are grouped together in Floating-point and integer control in the XL Fortran Compiler Reference, especially the -qfloat option. The XL Fortran compiler also provides three intrinsic modules for exception handling and IEEE arithmetic support to help you write IEEE module-compliant code that can be more portable. See IEEE Modules and Support in the XL Fortran Language Reference for details. The use of the compiler options for floating-point calculations affects the accuracy, performance, and possibly the correctness of floating-point calculations. Although the default values for the options were chosen to provide efficient and correct execution of most programs, you may need to specify nondefault options for your applications to work the way you want. We strongly advise you to read this section before using these options. Note: The discussions of single-, double-, and extended-precision calculations in this section all refer to the default situation, with -qrealsize=4 and no -qautodbl specified. If you change these settings, keep in mind that the size of a Fortran REAL, DOUBLE PRECISION, and so on may change, but single precision, double precision, and extended precision (in lowercase) still refer to 4-, 8-, and 16-byte entities respectively. IEEE floating-point overview The ANSI/IEEE Standard for Binary Floating-Point Arithmetic, ANSI/IEEE Std 754-1985 and IEEE Std 754-2008 and the details of how it applies to XL Fortran on specific hardware platforms, are summarized in the following topics. For information on the Fortran 2003 IEEE Module and arithmetic support, see the XL Fortran Language Reference. Compiling for strict IEEE conformance By default, XL Fortran follows most, but not all of the rules in the IEEE standard. To compile for strict compliance with the standard: v Use the compiler option -qfloat=nomaf. v If the program changes the rounding mode at run time, include rrm among the -qfloat suboptions. v If the data or program code contains signaling NaN values (NAN), include nans among the -qfloat suboptions. (A signaling NaN is different from a quiet NaN; you must explicitly code it into the program or data or create it by using the -qinitauto or -qinitalloc compiler option.) © Copyright IBM Corp. 1990, 2012 285 v If you are compiling with -O3, or a higher base optimization level, include the -qstrict option. You can also use the -qstrict suboptions to refine the level of control for the transformations performed by the optimizers. Related reference: See -qstrict in the Compiler Reference IEEE Single- and double-precision values XL Fortran encodes single-precision and double-precision values in IEEE format. For the range and representation, see Real in the XL Fortran Language Reference. IEEE extended-precision values The IEEE standard suggests, but does not mandate, a format for extended-precision values. XL Fortran does not use this format. “Extended-precision values” on page 289 describes the format that XL Fortran uses. Infinities and NaNs For single-precision real values: v Positive infinity is represented by the bit pattern X'7F80 0000'. v Negative infinity is represented by the bit pattern X'FF80 0000'. v A signaling NaN is represented by any bit pattern between X'7F80 0001' and X'7FBF FFFF' or between X'FF80 0001' and X'FFBF FFFF'. v A quiet NaN is represented by any bit pattern between X'7FC0 0000' and X'7FFF FFFF' or between X'FFC0 0000' and X'FFFF FFFF'. For double-precision real values: v Positive infinity is represented by the bit pattern X'7FF00000 00000000'. v Negative infinity is represented by the bit pattern X'FFF00000 00000000'. v A signaling NaN is represented by any bit pattern between X'7FF00000 00000001' and X'7FF7FFFF FFFFFFFF' or between X'FFF00000 00000001' and X'FFF7FFFF FFFFFFFF'. v A quiet NaN is represented by any bit pattern between X'7FF80000 00000000' and X'7FFFFFFF FFFFFFFF' or between X'FFF80000 00000000' and X'FFFFFFFF FFFFFFFF'. These values do not correspond to any Fortran real constants. You can generate all of these by encoding the bit pattern directly, or by using the ieee_value function provided in the ieee_arithmetic intrinsic module. Using the ieee_value function is the preferred programming technique, as it is allowed by the Fortran 2003 standard and the results are portable. Encoding the bit pattern directly could cause portability problems on machines using different bit patterns for the different values. All except signaling NaN values can occur as the result of arithmetic operations: 286 XL Fortran: Optimization and Programming Guide $ cat fp_values.f real plus_inf, minus_inf, plus_nanq, minus_nanq, nans real large data plus_inf /z’7f800000’/ data minus_inf /z’ff800000’/ data plus_nanq /z’7fc00000’/ data minus_nanq /z’ffc00000’/ data nans /z’7f800001’/ print *, ’Special values:’, plus_inf, minus_inf, plus_nanq, minus_nanq, nans ! They can also occur as the result of operations. large = 10.0 ** 200 print *, ’Number too big for a REAL:’, large * large print *, ’Number divided by zero:’, (-large) / 0.0 print *, ’Nonsensical results:’, plus_inf - plus_inf, sqrt(-large) ! To find if something is a NaN, compare it to itself. print *, ’Does a quiet NaN equal itself:’, plus_nanq .eq. plus_nanq print *, ’Does a signaling NaN equal itself:’, nans .eq. nans ! Only for a NaN is this comparison false. end $ bgxlf95 -o fp_values fp_values.f ** _main === End of Compilation 1 === 1501-510 Compilation successful for file fp_values.f. $ fp_values Special values: INF -INF NAN -NAN NAN Number too big for a REAL: INF Number divided by zero: -INF Nonsensical results: NAN NAN Does a quiet NaN equal itself: F Does a signaling NaN equal itself: F Exception-handling model The IEEE standard defines several exception conditions that can occur: OVERFLOW The exponent of a value is too large to be represented. UNDERFLOW A nonzero value is so small that it cannot be represented without an extraordinary loss of accuracy. The value can be represented only as zero or a subnormal number (denorm). ZERODIVIDE A finite nonzero value is divided by zero. INVALID Operations are performed on values for which the results are not defined. These include: v Operations on signaling NaN values v infinity - infinity v 0.0 * infinity v 0.0 / 0.0 v mod(x,y) or ieee_rem(x,y) (or other remainder functions) when x is infinite or y is zero v The square root of a negative number except -0.0 v Conversion of a floating-point number to an integer when the converted value cannot be represented faithfully Chapter 10. Implementation details of XL Fortran floating-point processing 287 v Comparisons involving NaN values INEXACT A computed value cannot be represented exactly, so a rounding error is introduced. (This exception is very common.) XL Fortran always detects these exceptions when they occur, but by default does not take any special action. Calculation continues, usually with a NaN or infinity value as the result. If you want to be automatically informed when an exception occurs, you can turn on exception trapping through compiler options or calls to intrinsic subprograms. However, different results, intended to be manipulated by exception handlers, are produced: Table 30. Results of IEEE exceptions, with and without trapping enabled Overflow Underflow Zerodivide Invalid Inexact Exceptions not INF Denormalized INF NaN Rounded result enabled (default) number Exceptions Unnormalized Unnormalized No result No result Rounded result enabled number with number with biased exponent biased exponent Note: Because different results are possible, it is very important to make sure that any exceptions that are generated are handled correctly. See “Detecting and trapping floating-point exceptions” on page 294 for instructions on doing so. Hardware-specific floating-point overview Single- and double-precision values and extended-precision values for hardware-specific floating-point processing are described in the following topics. Single- and double-precision values The PowerPC floating-point hardware performs calculations in either IEEE single-precision (equivalent to REAL(4) in Fortran programs) or IEEE double-precision (equivalent to REAL(8) in Fortran programs). Keep the following considerations in mind: v Double precision provides greater range (approximately 10**(-308) to 10**308) and precision (about 15 decimal digits) than single precision (approximate range 10**(-38) to 10**38, with about 7 decimal digits of precision). v Computations that mix single and double operands are performed in double precision, which requires conversion of the single-precision operands to double-precision. These conversions do not affect performance. v Double-precision values that are converted to single-precision (such as when you specify the SNGL intrinsic or when a double-precision computation result is stored into a single-precision variable) require rounding operations. A rounding operation produces the correct single-precision value, which is based on the IEEE rounding mode in effect. The value may be less precise than the original double-precision value, as a result of rounding error. Conversions from double-precision values to single-precision values may reduce the performance of your code. v Programs that manipulate large amounts of floating-point data may run faster if they use REAL(4) rather than REAL(8) variables. (You need to ensure that REAL(4) variables provide you with acceptable range and precision.) The 288 XL Fortran: Optimization and Programming Guide programs may run faster because the smaller data size reduces memory traffic, which can be a performance bottleneck for some applications. The floating-point hardware also provides a special set of double-precision operations that multiply two numbers and add a third number to the product. These combined multiply-add (MAF) operations are performed at the same speed at which either an individual multiply or add is performed. The MAF functions provide an extension to the IEEE 754-1985 standard (but are in the 754-2008 standard) because they perform the multiply and add with one (rather than two) rounding errors. The MAF functions are faster and more accurate than the equivalent separate operations. Extended-precision values XL Fortran extended precision is not in the format suggested by the IEEE standard, which suggests extended formats using more bits in both the exponent (for greater range) and the fraction (for greater precision). XL Fortran extended precision, equivalent to REAL(16) in Fortran programs, is implemented in software. Extended precision provides the same range as double precision (about 10**(-308) to 10**308) but more precision (a variable amount, about 31 decimal digits or more). The software support is restricted to round-to-nearest mode. Programs that use extended precision must ensure that this rounding mode is in effect when extended-precision calculations are performed. See “Selecting the rounding mode” on page 290 for the different ways you can control the rounding mode. Programs that specify extended-precision values as hexadecimal, octal, binary, or Hollerith constants must follow these conventions: v Extended-precision numbers are composed of two double-precision numbers with different magnitudes that do not overlap (except when the number is zero or close to zero). That is, the binary exponents differ by at least the number of fraction bits in a REAL(8). The high-order double-precision value (the one that comes first in storage) must have the larger magnitude. The value of the extended-precision number is the sum of the two double-precision values. v For a value of NaN or infinity, you must encode one of these values within the high-order double-precision value. The low-order value is not significant except that it cannot be set to NaN when the high-order value is infinity. Because an XL Fortran extended-precision value can be the sum of two values with greatly different exponents, leaving a number of assumed zeros in the fraction, the format actually has a variable precision with a minimum of about 31 decimal digits. You get more precision in cases where the exponents of the two double values differ in magnitude by more than the number of digits in a double-precision value. This encoding allows an efficient implementation intended for applications requiring more precision but no more range than double precision. Note: 1. In the discussions of rounding errors because of compile-time folding of expressions, keep in mind that this folding produces different results for extended-precision values more often than for other precisions. 2. Special numbers, such as NaN and infinity, are not fully supported for extended-precision values. Arithmetic operations do not necessarily propagate these numbers in extended precision. Chapter 10. Implementation details of XL Fortran floating-point processing 289 3. XL Fortran does not always detect floating-point exception conditions (see “Detecting and trapping floating-point exceptions” on page 294) for extended-precision values. If you turn on floating-point exception trapping in programs that use extended precision, XL Fortran may also generate signals in cases where an exception condition does not really occur. How XL Fortran rounds floating-point calculations Understanding rounding operations in XL Fortran can help you get predictable, consistent results. It can also help you make informed decisions when you have to make tradeoffs between speed and accuracy. In general, floating-point results from XL Fortran programs are more accurate than those from other implementations because of MAF operations and the higher precision used for intermediate results. If identical results are more important to you than the extra precision and performance of the XL Fortran defaults, read “Duplicating the floating-point results of other systems” on page 293. Selecting the rounding mode To change the rounding mode in a program, you can call the fpsets and fpgets routines, which use an array of logicals named fpstat, defined in the include files /opt/ibmcmp/xlf/bg/14.1/include/fpdt.h and /opt/ibmcmp/xlf/bg/14.1/include/ fpdc.h. The fpstat array elements correspond to the bits in the floating-point status and control register. For floating-point rounding control, the array elements fpstat(fprn1) and fpstat(fprn2) are set as specified in the following table: Table 31. Rounding-mode bits to use with fpsets and fpgets fpstat(fprn1) fpstat(fprn2) Rounding Mode Enabled .true. .true. Round towards -infinity. .true. .false. Round towards +infinity. .false. .true. Round towards zero. .false. .false. Round to nearest. For example: program fptest include ’fpdc.h’ call fpgets( fpstat ) ! Get current register values. if ( (fpstat(fprn1) .eqv. .false.) .and. + (fpstat(fprn2) .eqv. .false.)) then print *, ’Before test: Rounding mode is towards nearest’ print *, ’ 2.0 / 3.0 = ’, 2.0 / 3.0 print *, ’ -2.0 / 3.0 = ’, -2.0 / 3.0 end if call fpgets( fpstat ) ! Get current register values. fpstat(fprn1) = .TRUE. ! These 2 lines mean round towards fpstat(fprn2) = .FALSE. ! +infinity. call fpsets( fpstat ) r = 2.0 / 3.0 print *, ’Round towards +infinity: 2.0 / 3.0= ’, r call fpgets( fpstat ) ! Get current register values. fpstat(fprn1) = .TRUE. ! These 2 lines mean round towards 290 XL Fortran: Optimization and Programming Guide fpstat(fprn2) = .TRUE. ! -infinity. call fpsets( fpstat ) r = -2.0 / 3.0 print *, ’Round towards -infinity: -2.0 / 3.0= ’, r end ! This block data program unit initializes the fpstat array, and so on. block data include ’fpdc.h’ include ’fpdt.h’ end XL Fortran also provides several procedures that allow you to control the floating-point status and control register of the processor directly. These procedures are more efficient than the fpsets and fpgets subroutines because they are mapped into inlined machine instructions that manipulate the floating-point status and control register (fpscr) directly. XL Fortran supplies the get_round_mode() and set_round_mode() procedures in the xlf_fp_util module. These procedures return and set the current floating-point rounding mode, respectively. For example: program fptest use, intrinsic :: xlf_fp_util integer(fpscr_kind) old_fpscr if ( get_round_mode() == fp_rnd_rn ) then print *, ’Before test: Rounding mode is towards nearest’ print *, ’ 2.0 / 3.0 = ’, 2.0 / 3.0 print *, ’ -2.0 / 3.0 = ’, -2.0 / 3.0 end if old_fpscr = set_round_mode( fp_rnd_rp ) r = 2.0 / 3.0 print *, ’Round towards +infinity: 2.0 / 3.0 = ’, r old_fpscr = set_round_mode( fp_rnd_rm ) r = -2.0 / 3.0 print *, ’Round towards -infinity: -2.0 / 3.0 = ’, r end XL Fortran supplies the ieee_get_rounding_mode() and ieee_set_rounding_mode() procedures in the ieee_arithmetic module. These portable procedures retrieve and set the current floating-point rounding mode, respectively. For example: program fptest use, intrinsic :: ieee_arithmetic type(ieee_round_type) current_mode call ieee_get_rounding_mode( current_mode ) if ( current_mode == ieee_nearest ) then print *, ’Before test: Rounding mode is towards nearest’ print *, ’ 2.0 / 3.0 = ’, 2.0 / 3.0 print *, ’ -2.0 / 3.0 = ’, -2.0 / 3.0 end if call ieee_set_rounding_mode( ieee_up ) r = 2.0 / 3.0 print *, ’Round towards +infinity: 2.0 / 3.0 = ’, r call ieee_set_rounding_mode( ieee_down ) r = -2.0 / 3.0 print *, ’Round towards -infinity: -2.0 / 3.0 = ’, r end Chapter 10. Implementation details of XL Fortran floating-point processing 291 Notes: 1. Extended-precision floating-point values must only be used in round-to-nearest mode. 2. For thread-safety and reentrancy, the include file /opt/ibmcmp/xlf/bg/14.1/ include/fpdc.h contains a THREADLOCAL directive that is protected by the trigger constant IBMT. The invocation commands bgxlf_r, bgxlf90_r, bgxlf95_r, bgxlf2003_r, and bgxlf2008_r turn on the -qthreaded compiler option by default, which in turn implies the trigger constant IBMT. If you are including the file /opt/ibmcmp/xlf/bg/14.1/include/fpdc.h in code that is not intended to be threadsafe, do not specify IBMT as a trigger constant. 3. Compile a program that changes the rounding mode with -qfloat=rrm. Minimizing rounding errors There are several strategies for handling rounding errors and other unexpected, slight differences in calculated results. You may want to consider one or more of the following strategies: v Minimizing the amount of overall rounding v Delaying as much rounding as possible to run time v Ensuring that if some rounding is performed in a mode other than round-to-nearest, all rounding is performed in the same mode Minimizing overall rounding Rounding operations, especially in loops, reduce code performance and may have a negative effect on the precision of computations. Consider using double-precision variables instead of single-precision variables when you store the temporary results of double-precision calculations, and delay rounding operations until the final result is computed. Delaying rounding until run time The compiler evaluates floating-point expressions during compilation when it can, so that the resulting program does not run more slowly due to unnecessary runtime calculations. However, the results of the compiler's evaluation might not match exactly the results of the runtime calculation. To delay these calculations until run time, specify the nofold suboption of the -qfloat option. The results may still not be identical; for example, calculations in DATA and PARAMETER statements are still performed at compile time. The differences in results due to fold or nofold are greatest for programs that perform extended-precision calculations or are compiled with the -O option or both. Ensuring that the rounding mode is consistent You can change the rounding mode from its default setting of round-to-nearest. (See for examples.) If you do so, you must be careful that all rounding operations for the program use the same mode: v Specify the equivalent setting on the -qieee option, so that any compile-time calculations use the same rounding mode. v Specify the rrm suboption of the -qfloat option, so that the compiler does not perform any optimizations that require round-to-nearest rounding mode to work correctly. 292 XL Fortran: Optimization and Programming Guide For example, you might compile a program like the one in “Selecting the rounding mode” on page 290 with this command if the program consistently uses round-to-plus-infinity mode: bgxlf95 -qieee=plus -qfloat=rrm changes_rounding_mode.f Duplicating the floating-point results of other systems To duplicate the double-precision results of programs on systems with different floating-point architectures (without multiply-add instructions), specify the nomaf suboption of the -qfloat option. This suboption prevents the compiler from generating any multiply-add instructions. This results in decreased accuracy and performance but provides strict conformance to the IEEE standard for double-precision arithmetic. To duplicate the results of programs where the default size of REAL items is different from that on systems running XL Fortran, use the -qrealsize option to change the default REAL size when compiling with XL Fortran. If the system whose results you want to duplicate preserves full double precision for default real constants that are assigned to DOUBLE PRECISION variables, use the -qdpc or -qrealsize option. If results consistent with other systems are important to you, include norsqrt and nofold in the settings for the -qfloat option. If you specify the option -O3, -O4,or -O5, include -qstrict and any necessary suboptions too. Related information: See -qarch in the Compiler Reference See -qfloat in the Compiler Reference See -qrealsize in the Compiler Reference See -qstrict in the Compiler Reference Maximizing floating-point performance If performance is your primary concern and you want your program to be relatively safe but do not mind if results are slightly different (generally more precise) from what they would be otherwise, optimize the program with the -O option, and specify -qfloat=rsqrt:hssngl:fltint. The following topics describe the functions of these suboptions: v The rsqrt suboption replaces division by a square root with multiplication by the reciprocal of the root, a faster operation that may not produce precisely the same result. v The hssngl suboption improves the performance of single-precision (REAL(4)) floating-point calculations by suppressing rounding operations that are required by the Fortran language but are not necessary for correct program execution. The results of floating-point expressions are kept in double precision where the original program would round them to single-precision. These results are then used in some later expressions instead of the rounded results. To detect single-precision floating-point overflows and underflows, rounding operations are still inserted when double-precision results are stored into single-precision memory locations. However, if optimization removes such a Chapter 10. Implementation details of XL Fortran floating-point processing 293 store operation, hssngl also removes the corresponding rounding operation, possibly preventing the exception. (Depending on the characteristics of your program, you may or may not care whether the exception happens.) The hssngl suboption is safe for all types of programs because it always only increases the precision of floating-point calculations. Program results may differ because of the increased precision and because of avoidance of some exceptions. v The fltint suboption speeds up float-to-integer conversions by reducing error checking for overflows when the program is compiled to run on older processors. You should make sure that any floats that are converted to integers are not outside the range of the corresponding integer types. Detecting and trapping floating-point exceptions The IEEE standard for floating-point arithmetic defines a number of exception (or error) conditions that might require special care to avoid or recover from. The following topics are intended to help you make your programs work safely in the presence of such exception conditions while sacrificing the minimum amount of performance. The floating-point hardware always detects a number of floating-point exception conditions (which the IEEE standard rigorously defines): overflow, underflow, zerodivide, invalid, and inexact. By default, the only action that occurs is that a status flag is set. The program continues without a problem (although the results from that point on may not be what you expect). If you want to know when an exception occurs, you can arrange for one or more of these exception conditions to generate a signal. The signal causes a branch to a handler routine. The handler receives information about the type of signal and the state of the program when the signal occurred. It can produce a core dump, display a listing showing where the exception occurred, modify the results of the calculation, or carry out some other processing that you specify. The XL Fortran compiler uses the operating system facilities for working with floating-point exception conditions. These facilities indicate the presence of floating-point exceptions by generating SIGFPE signals. Compiler features for trapping floating-point exceptions To turn on XL Fortran exception trapping, compile the program with the -qflttrap option and some combination of suboptions that includes enable. This option uses trap operations to detect floating-point exceptions and generates SIGFPE signals when exceptions occur, provided that a signal handler for SIGFPE is installed. -qflttrap also has suboptions that correspond to the names of the exception conditions. In particular, use the -qflttrap=qpxstore suboption to detect NaN or infinity values in QPX vectors. The compiler generates stores with indicating instructions for QPX vectors in registers. For example, if you are concerned with handling overflow, underflow, and QPX floating point exceptions, you can specify the following command: bgxlf -qflttrap=overflow:underflow:qpxstore:enable compute_pi.f You only need enable when you are compiling the main program. However, it is very important and does not cause any problems if you specify it for other files, so always include it when you use -qflttrap. 294 XL Fortran: Optimization and Programming Guide An advantage of this approach is that performance impact is relatively low. However, this approach only traps exceptions that occur in code that you compiled with -qflttrap, which does not include system library routines. Notes: 1. If your program depends on floating-point exceptions occurring for particular operations, also specify -qfloat suboptions that include nofold. Otherwise, the compiler might replace an exception-producing calculation with a constant NaN or infinity value, or it might eliminate an overflow in a single-precision operation. 2. The suboptions of the -qflttrap option replace an earlier technique that required you to modify your code with calls to the fpsets and fpgets procedures. You no longer require these calls for exception handling if you use the appropriate -qflttrap settings. Attention: If your code contains fpsets calls that enable checking for floating-point exceptions and you do not use the -qflttrap option when compiling the whole program, the program will produce unexpected results if exceptions occur, as explained in Table 30 on page 288. Installing an exception handler When a program that uses the XL Fortran or Blue Gene/Q exception-detection facilities encounters an exception condition, it receives a signal from the operating system. This causes a branch to whatever handler is specified by the program. By default, programs on Blue Gene/Q do not trap on floating-point exceptions unless a signal handler is installed. To produce a core file, you can use the xl__trcedump signal handler described below. If you want to install a SIGTRAP or SIGFPE signal handler, use the -qsigtrap option. It allows you to specify an XL Fortran handler that produces a traceback or to specify a handler you have written: bgxlf95 -qflttrap=ov:und:en pi.f # Dump core on an exception bgxlf95 -qflttrap=ov:und:en -qsigtrap pi.f # Uses the xl__trce handler bgxlf95 -qflttrap=ov:und:en -qsigtrap=return_22_over_7 pi.f # Uses any other handler You can also install an alternative exception handler, either one supplied by XL Fortran or one you have written yourself, by calling the SIGNAL subroutine (defined in /opt/ibmcmp/xlf/bg/14.1/include/fexcp.h): INCLUDE ’fexcp.h’ CALL SIGNAL(SIGTRAP,handler_name) CALL SIGNAL(SIGFPE,handler_name) The XL Fortran exception handlers and related routines are: xl__ieee Produces a traceback and an explanation of the signal and continues execution by supplying the default IEEE result for the failed computation. This handler allows the program to produce the same results as if exception detection was not turned on. xl__trce Produces a traceback and stops the program. xl__trcedump Produces a traceback and a core file and stops the program. xl__sigdump Provides a traceback that starts from the point at which it is called and provides information about the signal. You can only call it from inside a Chapter 10. Implementation details of XL Fortran floating-point processing 295 user-written signal handler. It does not stop the program. To successfully continue, the signal handler must perform some cleanup after calling this subprogram. xl__trbk Provides a traceback that starts from the point at which it is called. You call it as a subroutine from your code, rather than specifying it with the -qsigtrap option. It requires no parameters. It does not stop the program. All of these handler names contain double underscores to avoid duplicating names that you declared in your program. All of these routines work for both SIGTRAP and SIGFPE signals. You can use the -g compiler option to get line numbers in the traceback listings. The file /opt/ibmcmp/xlf/bg/14.1/include/fsignal.h defines a Fortran derived type similar to the ucontext_t structure in /usr/include/sys/ucontext.h system header. You can write a Fortran signal handler that accesses this derived type. “Sample programs for exception handling” on page 299 lists some sample programs that illustrate how to use these signal handlers or write your own. Also see the SIGNAL procedure in the XL Fortran Language Reference for more information. Producing a core file To produce a core file, specify the xl__trcedump handler. Controlling the floating-point status and control register Before the introduction of -qflttrap suboptions or the -qsigtrap options, most of the processing for floating-point exceptions required you to change your source files to turn on exception trapping or install a signal handler. Although you can still do so, for any new applications, we recommend that you use the options instead. To control exception handling at run time, compile without the enable suboption of the -qflttrap option: bgxlf95 -qflttrap compute_pi.f # Check all exceptions, but do not trap. bgxlf95 -qflttrap=ov compute_pi.f # Check one type, but do not trap. Then, inside your program, manipulate the fpstats array (defined in the include file /opt/ibmcmp/xlf/bg/14.1/include/fpdc.h) and call the fpsets subroutine to specify which exceptions should generate traps. See the sample program that uses fpsets and fpgets in “Selecting the rounding mode” on page 290. Another method is to use the set_fpscr_flags() subroutine in the xlf_fp_util module. This subroutine allows you to set the floating-point status and control register flags you specify in the MASK argument. Flags that you do not specify in MASK remain unaffected. MASK must be of type INTEGER(FPSCR_KIND). For example: USE, INTRINSIC :: xlf_fp_util INTEGER(FPSCR_KIND) SAVED_FPSCR INTEGER(FP_MODE_KIND) FP_MODE SAVED_FPSCR = get_fpscr() ! Saves the current value of ! the fpscr register. 296 XL Fortran: Optimization and Programming Guide CALL set_fpscr_flags(TRP_DIV_BY_ZERO) ! Enables trapping of ! ... ! divide-by-zero. SAVED_FPSCR=set_fpscr(SAVED_FPSCR) ! Restores fpscr register. Another method is to use the ieee_set_halting_mode subroutine in the ieee_exceptions module. This portable subroutine allows you to set the halting (trapping) status for any FPSCR exception flags. For example: USE, INTRINSIC :: ieee_exceptions TYPE(IEEE_STATUS_TYPE) SAVED_FPSCR CALL ieee_get_status(SAVED_FPSCR) ! Saves the current value of the ! fpscr register CALL ieee_set_halting_mode(IEEE_DIVIDE_BY_ZERO, .TRUE.) ! Enabled trapping ! ... ! of divide-by-zero. CALL IEEE_SET_STATUS(SAVED_FPSCR) ! Restore fpscr register xlf_fp_util procedures The xlf_fp_util procedures allow you to query and control the floating-point status and control register (fpscr) of the processor directly. These procedures are more efficient than the fpsets and fpgets subroutines because they are mapped into inlined machine instructions that manipulate the floating-point status and control register directly. The intrinsic module, xlf_fp_util, contains the interfaces and data type definitions for these procedures and the definitions for the named constants that are needed by the procedures. This module enables type checking of these procedures at compile time rather than link time. The following files are supplied for the xlf_fp_util module: File names File type Locations xlf_fp_util.mod module symbol file /opt/ibmcmp/xlf/bg/14.1/include64 (64-bit) To use the procedures, you must add a USE XLF_FP_UTIL statement to your source file. For more information, see the USE statement in the XL Fortran Language Reference. When compiling with the -U option, you must code the names of these procedures in all lowercase. For a list of the xlf_fp_util procedures, see the Service and utility procedures section in the XL Fortran Language Reference. fpgets and fpsets subroutines The fpsets and fpgets subroutines provide a way to manipulate or query the floating-point status and control register. Instead of calling the operating system routines directly, you pass information back and forth in fpstat, an array of logicals. The following table shows the most commonly used array elements that deal with exceptions: Chapter 10. Implementation details of XL Fortran floating-point processing 297 Table 32. Exception bits to use with fpsets and fpgets Array Element to Array Element to Check if Exception Set to Enable Occurred Exception Indicated When .TRUE. n/a fpstat(fpfx) Floating-point exception summary n/a fpstat(fpfex) Floating-point enabled exception summary fpstat(fpve) fpstat(fpvx) Floating-point invalid operation exception summary fpstat(fpoe) fpstat(fpox) Floating-point overflow exception fpstat(fpue) fpstat(fpux) Floating-point underflow exception fpstat(fpze) fpstat(fpzx) Zero-divide exception fpstat(fpxe) fpstat(fpxx) Inexact exception fpstat(fpve) fpstat(fpvxsnan) Floating-point invalid operation exception (signaling NaN) fpstat(fpve) fpstat(fpvxisi) Floating-point invalid operation exception (INF-INF) fpstat(fpve) fpstat(fpvxidi) Floating-point invalid operation exception (INF/INF) fpstat(fpve) fpstat(fpvxzdz) Floating-point invalid operation exception (0/0) fpstat(fpve) fpstat(fpvximz) Floating-point invalid operation exception (INF*0) fpstat(fpve) fpstat(fpvxvc) Floating-point invalid operation exception (invalid compare) n/a fpstat(fpvxsoft) Floating-point invalid operation exception (software request), PowerPC only n/a fpstat(fpvxsqrt) Floating-point invalid operation exception (invalid square root), PowerPC only n/a fpstat(fpvxcvi) Floating-point invalid operation exception (invalid integer convert), PowerPC only To explicitly check for specific exceptions at particular points in a program, use fpgets and then test whether the elements in fpstat have changed. Once an exception has occurred, the corresponding exception bit (second column in the preceding table) is set until it is explicitly reset, except for fpstat(fpfx), fpstat(fpvx), and fpstat(fpfex), which are reset only when the specific exception bits are reset. An advantage of using the fpgets and fpsets subroutines (as opposed to controlling everything with suboptions of the -qflttrap option) includes control over granularity of exception checking. For example, you might only want to test if an exception occurred anywhere in the program when the program ends. The disadvantages of this approach include the following: v You have to change your source code. v These routines differ from what you may be accustomed to on other platforms. 298 XL Fortran: Optimization and Programming Guide For example, to trap floating-point overflow exceptions but only in a certain section of the program, you would set fpstat(fpoe) to .TRUE. and call fpsets. After the exception occurs, the corresponding exception bit, fpstat(fpox),is .TRUE. until the program runs: call fpgets(fpstat) fpstat(fpox) = .FALSE. call fpsets(fpstat) ! resetting fpstat(fpox) to .FALSE. Sample programs for exception handling Sample programs contained in /opt/ibmcmp/xlf/bg/14.1/samples/floating_point illustrate different aspects of exception handling: flttrap_handler.c and flttrap_test.f A sample exception handler that is written in C and a Fortran program that uses it. xl__ieee.F and xl__ieee.c Exception handlers that are written in Fortran and C that show how to substitute particular values for operations that produce exceptions. Even when you use support code such as this, the implementation of XL Fortran exception handling does not fully support the exception-handling environment that is suggested by the IEEE floating-point standard. check_fpscr.f and postmortem.f Show how to work with the fpsets and fpgets procedures and the fpstats array. fhandler.F Shows a sample Fortran signal handler and demonstrates the xl__sigdump procedure. xl__trbk_test.f Shows how to use the xl__trbk procedure to generate a traceback listing without stopping the program. The sample programs are strictly for illustrative purposes only. Causing exceptions for particular variables To mark a variable as “do not use”, you can encode a special value called a signaling NaN in it. This causes an invalid exception condition any time that variable is used in a calculation. If you use this technique, use the nans suboption of the -qfloat option, so that the program properly detects all cases where a signaling NaN is used, and one of the methods already described to generate corresponding SIGFPE signals. Minimizing the performance impact of floating-point exception trapping If you need to deal with floating-point exception conditions but are concerned that doing so will make your program too slow, here are some techniques that can help minimize the performance impact: v Consider using only a subset of the overflow, underflow, zerodivide, invalid, and inexact suboptions with the -qflttrap option if you can identify some conditions that will never happen or you do not care about. In particular, because an inexact exception occurs for each rounding error, you probably should not check for it if performance is important. Chapter 10. Implementation details of XL Fortran floating-point processing 299 300 XL Fortran: Optimization and Programming Guide Chapter 11. Porting programs to XL Fortran XL Fortran provides many features intended to make it easier to take programs that were originally written for other computer systems or compilers and recompile them with XL Fortran. Outline of the porting process The process of porting a typical program is described in this topic. The process for porting a typical program looks like this: 1. Identify any nonportable language extensions or subroutines that you used in the original program. Check to see if any of them are supported by XL Fortran: v Language extensions are identified in the XL Fortran Language Reference. v Some extensions require you to specify an XL Fortran compiler option; you can find these options listed in the Portability and migration options table in the XL Fortran Compiler Reference. 2. For any nonportable features that XL Fortran does not support, modify the source files to remove or work around them. 3. Do the same for any implementation-dependent features. For example, if your program relies on exact bit-pattern representation of floating-point values or uses system-specific file names, you may need to change it. 4. Compile the program with XL Fortran. If any compilation problems occur, fix them and recompile and fix any additional errors until the program compiles successfully. 5. Run the XL Fortran-compiled program and compare the output with the output from the other system. If the results are substantially different, there are probably still some implementation-specific features that need to be changed. If the results are only marginally different (for example, if XL Fortran produces a different number of digits of precision or a number differs in the last decimal place), decide whether the difference is significant enough to investigate further. You may be able to fix these differences. Before porting programs to XL Fortran, read the tips in the following sections so that you know in advance what compatibility features XL Fortran offers. Portability of directives XL Fortran supports many directives available with other Fortran products. This ensures easy portability between products. If your code contains trigger_constants other than the defaults in XL Fortran, you can use the -qdirective compiler option to specify them. For instance, if you are porting CRAY code contained in a file xx.f, you would use the following command to add the CRAY trigger_constant: bgxlf95 xx.f -qdirective=mic\$ For fixed source form code, in addition to the ! value for the trigger_head portion of the directive, XL Fortran also supports the trigger_head values C, c, and *. © Copyright IBM Corp. 1990, 2012 301 For more information, see the -qdirective option in the XL Fortran Compiler Reference. XL Fortran supports a number of programming terms as synonyms to ease the effort of porting code from other Fortran products. Those terms that are supported are dependent on context, as indicated in the following tables: Table 33. PARALLEL DO Clauses and their XL Fortran synonyms PARALLEL DO Clause XL Fortran Synonym LASTLOCAL LASTPRIVATE LOCAL PRIVATE SCHEDULE MP_SCHEDTYPE and CHUNK SAVELAST LASTPRIVATE SHARE SHARED NEW PRIVATE Table 34. PARALLEL DO scheduling types and their XL Fortran synonyms Scheduling Type XL Fortran Synonym GSS GUIDED INTERLEAVE STATIC(1) INTERLEAVED STATIC(1) INTERLEAVE(n) STATIC(n) INTERLEAVED(n) STATIC(n) SIMPLE STATIC Table 35. PARALLEL SECTIONS clauses and their XL Fortran synonyms PARALLEL SECTIONS Clause XL Fortran Synonym LOCAL PRIVATE SHARE SHARED NEW PRIVATE Common industry extensions that XL Fortran supports XL Fortran allows many of the same FORTRAN 77 extensions as other popular compilers. These extensions include: Refer to XL Fortran Language Reference Extension Section(s) Typeless constants Typeless literal constants *len length specifiers for types Data types BYTE data type Byte 302 XL Fortran: Optimization and Programming Guide Refer to XL Fortran Language Reference Extension Section(s) Long variable names Names Lower case Names Mixing integers and logicals (with -qintlog option) Evaluation of expressions Character-count Q edit descriptor (with -qqcount option) Q (Character Count) Editing Intrinsics for counting set bits in registers and determining POPCNT, POPPAR data-object parity 64-bit data types (INTEGER(8), REAL(8), COMPLEX(8), and Integer Real Complex LOGICAL(8)), including support for default 64-bit types (with Logical -qintsize and -qrealsize options) Integer POINTERs, similar to those supported by CRAY and Sun POINTER(integer) compilers. (XL Fortran integer pointer arithmetic uses increments of one byte, while the increment on CRAY computers is eight bytes. You may need to multiply pointer increments and decrements by eight to make programs ported from CRAY computers work properly.) Conditional vector merge (CVMGx) intrinsic functions CVMGx (TSOURCE, FSOURCE, MASK) Date and time service and utility functions (rtc, irtc, jdate, clock_, Service and utility timef, and date) procedures STRUCTURE, UNION, and MAP constructs Structure components, Union and map Finding nonstandard extensions XL Fortran supports a number of extensions to various language standards. Many of these extensions are so common that you need to keep in mind, when you port programs to other systems, that not all compilers have them. To find such extensions in your XL Fortran programs before beginning a porting effort, use the -qlanglvl option: $ # -qnoobject stops the compiler after parsing all the source, $ # giving a fast way to check for errors. $ # Look for anything above the base F77 standard. $ bgxlf -qnoobject -qlanglvl=77std f77prog.f ... $ # Look for anything above the F90 standard. $ bgxlf90 -qnoobject -qlanglvl=90std use_in_2000.f ... $ # Look for anything above the F95 standard. $ bgxlf95 -qnoobject -qlanglvl=95std use_in_2000.f ... Related reference: See -langlvl in the Compiler Reference See -qport in the Compiler Reference Mixing data types in statements The -qctyplss option lets you use character constant expressions in the same places that you use typeless constants. The -qintlog option lets you use integer Chapter 11. Porting programs to XL Fortran 303 expressions where you can use logicals, and vice versa. A kind type parameter must not be replaced with a logical constant even if -qintlog is on, nor by a character constant even if -qctyplss is on, nor can it be a typeless constant. Date and time routines Date and time routines, such as dtime, etime, and jdate, are accessible as Fortran subroutines. Other libc routines A number of other popular routines from the libc library, such as flush, getenv, and system, are also accessible as Fortran subroutines. Changing the default sizes of data types For porting from machines with larger or smaller word sizes, the -qintsize option lets you specify the default size for integers and logicals.The -qrealsize option lets you specify the default size for reals and complex components. Name conflicts between your procedures and XL Fortran intrinsic procedures If you have procedures with the same names as any XL Fortran intrinsic procedures, the program calls the intrinsic procedure. This situation is more likely with the addition of the many new Fortran 90, Fortran 95, Fortran 2003, and Fortran 2008 intrinsic procedures. If you still want to call your procedure, add explicit interfaces, EXTERNAL statements, or PROCEDURE statements for any procedures with conflicting names, or use the -qextern option when compiling. Reproducing results from other systems XL Fortran provides settings through the -qfloat option that help make floating-point results consistent with those from other IEEE systems; this subject is discussed in “Duplicating the floating-point results of other systems” on page 293. 304 XL Fortran: Optimization and Programming Guide Chapter 12. Sample Fortran programs The programs in the topics referenced here are provided as coding examples for XL Fortran. A number of these samples illustrate various aspects of SMP programming that may be new to many users. If you are new to SMP programming, you should examine these samples to gain a better understanding of the SMP coding style. You can compile and execute the first program to verify that the compiler is installed correctly and your user ID is set up to execute Fortran programs. Example1-XLFortran source file This is an example of an XL Fortran source file PROGRAM CALCULATE ! ! Program to calculate the sum of up to n values of x**3 ! where negative values are ignored. ! IMPLICIT NONE INTEGER I,N REAL SUM,X,Y READ(*,*) N WRITE(*,*) N SUM=0 DO I=1,N READ(*,*) X WRITE(*,*) X IF (X.GE.0.0) THEN Y=X**3 SUM=SUM+Y END IF END DO WRITE(*,*) ’This is the sum of the positive cubes:’,SUM END Execution results Running the program yields the following results: $ a.out 5 37 22 -4 19 6 This is the sum of the positive cubes: 68376.00000 Example 2 - valid C routine source file This is an example of a valid C routine source file used to execute Fortran test subroutines. /* * ******************************************************************** * This is a main function that creates threads to execute the Fortran * test subroutines. © Copyright IBM Corp. 1990, 2012 305 * ******************************************************************** */ #include extern char *optarg; extern int optind; static char *prog_name; #define MAX_NUM_THREADS 100 void *f_mt_exec(void *); void f_pre_mt_exec(void); void f_post_mt_exec(int *); void usage(void) { fprintf(stderr, "Usage: %s -t number_of_threads.\n", prog_name); exit(-1); } main(int argc, char *argv[]) { int i, c, rc; int num_of_threads, n[MAX_NUM_THREADS]; char *num_of_threads_p; pthread_attr_t attr; pthread_t tid[MAX_NUM_THREADS]; prog_name = argv[0]; while ((c = getopt(argc, argv, "t")) != EOF) { switch (c) { case ’t’: break; default: usage(); break; } } argc -= optind; argv += optind; if (argc < 1) { usage(); } num_of_threads_p = argv[0]; if ((num_of_threads = atoi(num_of_threads_p)) == 0) { fprintf(stderr, "%s: Invalid number of threads to be created <\n", prog_name, num_of_threads_p); exit(1); } else if (num_of_threads > MAX_NUM_THREADS) { fprintf(stderr, "%s: Cannot create more than 100 threads.\n", prog_name); exit(1); } 306 XL Fortran: Optimization and Programming Guide pthread_attr_init(&attr); pthread_attr_setdetachstate(&attr, PTHREAD_CREATE_JOINABLE); /* **************************************************************** * Execute the Fortran subroutine that prepares for multi-threaded * execution. * **************************************************************** */ f_pre_mt_exec(); for (i = 0; i < num_of_threads; i++) { n[i] = i; rc = pthread_create(&tid[i], &attr, f_mt_exec, (void *)&n[i]); if (rc != 0) { fprintf(stderr, "Failed to create thread %d.\n", i); exit(1); } } /* The attribute is no longer needed after threads are created. */ pthread_attr_destroy(&attr); for (i = 0; i < num_of_threads; i++) { rc = pthread_join(tid[i], NULL); if (rc != 0) { fprintf(stderr, "Failed to join thread %d. \n", i); } } /* * Execute the Fortran subroutine that does the check after * multi-threaded execution. */ f_post_mt_exec(&num_of_threads); exit(0); } ! *********************************************************************** ! This test case tests the writing list-directed to a single external ! file by many threads. ! *********************************************************************** subroutine f_pre_mt_exec() integer array(1000) common /x/ array doi=1,1000 array(i) = i end do open(10, file="fun10.out", form="formatted", status="replace") end subroutine f_post_mt_exec(number_of_threads) integer array(1000), array1(1000) common /x/ array close(10) open(10, file="fun10.out", form="formatted") doj=1,number_of_threads read(10, *) array1 doi=1,1000 Chapter 12. Sample Fortran programs 307 if (array1(i) /= array(i)) then print *, "Result is wrong." stop endif end do end do close(10, status="delete") print *, "Normal ending." end subroutine f_mt_exec(thread_number) integer thread_number integer array(1000) common /x/ array write(10, *) array end Example 3 - valid Fortran SMP source file This is an example of a valid Fortran SMP source file used to calculate the value of pi. !***************************************************************** !* This example uses a PARALLEL construct and a DO construct * !* to calculate the value of pi. * !***************************************************************** program compute_pi integer n, i real*8 w, x, pi, f, a f(a) = 4.d0 /(1.d0 + a*a) !! function to integrate pi = 0.0d0 !$OMP PARALLEL private(x, w, n), shared(pi) n = 10000 !! number of intervals w = 1.0d0/n !! calculate the interval size !$OMP DO reduction(+: pi) doi=1,n x=w*(i-0.5d0) pi = pi + f(x) enddo !$OMP END DO !$OMP END PARALLEL print *, "Computed pi = ", pi end Example 4 - invalid Fortran SMP source file This is an example of an invalid Fortran SMP source file. !***************************************************************** !* In this example, fort_sub is invoked by multiple threads. * !* * !* This example is not valid because * !* fort_sub and another_sub both declare /block/ to be * !* THREADPRIVATE. They intend to share the common block, but * !* they are executed via different threads. * !* * !* To "fix" this problem, one of the following approaches can * !* be taken: * !* (1) The code for another_sub should be brought into the loop.* !* (2) "j" should be passed as an argument to another_sub, and * !* the declaration for /block/ should be removed from * !* another_sub. * !* (3) The loop should be marked as "do not parallelize" by * !* using the directive "!$OMP PARALLEL DO IF(.FALSE.)". * 308 XL Fortran: Optimization and Programming Guide !***************************************************************** subroutine fort_sub() common /block/ j integer :: j !$OMP THREADPRIVATE(/block/) ! Each thread executing fort_sub ! obtains its own copy of /block/. integer a(10) ... !$OMP PARALLEL DO do index = 1,10 call another_sub(a(i)) enddo ... end subroutine fort_sub subroutine another_sub(aa) ! Multiple threads are used to integer aa ! execute another_sub. common /block/ j ! Each thread obtains a new copy integer :: j ! of the common block /block/. !$OMP THREADPRIVATE(/block/) aa = j ! The value of "j" is undefined. end subroutine another_sub Programming examples using the Pthreads library module These examples demonstrate the use of the Pthreads library module. !****************************************************************** !* Example 5 : Create a thread with Round_Robin scheduling policy.* !* For simplicity, we do not show any codes for error checking, * !* which would be necessary in a real program. * !****************************************************************** use, intrinsic::f_pthread integer(4) ret_val type(f_pthread_attr_t) attr type(f_pthread_t) thr ret_val = f_pthread_attr_init(attr) ret_val = f_pthread_attr_setschedpolicy(attr, SCHED_RR) ret_val = f_pthread_attr_setinheritsched(attr, PTHREAD_EXPLICIT_SCHED) ret_val = f_pthread_create(thr, attr, FLAG_DEFAULT, ent, integer_arg) ret_val = f_pthread_attr_destroy(attr) ...... Before you can manipulate a pthread attribute object, you need to create and initialize it. The appropriate interfaces must be called to manipulate the attribute objects. A call to f_pthread_attr_setschedpolicy sets the scheduling policy attribute to Round_Robin. Note that this does not affect newly created threads that inherit the scheduling property from the creating thread. For these threads, we explicitly call f_pthread_attr_setinheritsched to override the default inheritance attribute. The rest of the code is self-explanatory. !***************************************************************** !* Example 6 : Thread safety * !* In this example, we show that thread safety can be achieved * !* by using the push-pop cleanup stack for each thread. We * !* assume that the thread is in deferred cancellability-enabled * !* state. This means that any thread-cancel requests will be * !* put on hold until a cancellation point is encountered. * !* Note that f_pthread_cond_wait provides a * !* cancellation point. * Chapter 12. Sample Fortran programs 309 !***************************************************************** use, intrinsic::f_pthread integer(4) ret_val type(f_pthread_mutex_t) mutex type(f_pthread_cond_t) cond pointer(p, byte) ! Initialize mutex and condition variables before using them. ! For global variables this should be done in a module, so that they ! can be used by all threads. If they are local, other threads ! will not see them. Furthermore, they must be managed carefully ! (for example, destroy them before returning, to avoid dangling and ! undefined objects). mutex = PTHREAD_MUTEX_INITIALIZER cond = PTHREAD_COND_INITIALIZER ...... ! Doing something ...... ! This thread needs to allocate some memory area used to ! synchronize with other threads. However, when it waits on a ! condition variable, this thread may be canceled by another ! thread. The allocated memory may be lost if no measures are ! taken in advance. This will cause memory leakage. ret_val = f_pthread_mutex_lock(mutex) p = malloc(%val(4096)) ! Check condition. If it is not true, wait for it. ! This should be a loop. ! Since memory has been allocated, cleanup must be registered ! for safety during condition waiting. ret_val = f_pthread_cleanup_push(mycleanup, FLAG_DEFAULT, p) ret_val = f_pthread_cond_wait(cond, mutex) ! If this thread returns from condition waiting, the cleanup ! should be de-registered. call f_pthread_cleanup_pop(0) ! not execute ret_val = f_pthread_mutex_unlock(mutex) ! This thread will take care of p for the rest of its life...... ! mycleanup looks like: subroutine mycleanup(passed_in) pointer(passed_in, byte) external free call free(%val(passed_in)) end subroutine mycleanup 310 XL Fortran: Optimization and Programming Guide Notices This information was developed for products and services offered in the U.S.A. IBM may not offer the products, services, or features discussed in this document in other countries. Consult your local IBM representative for information on the products and services currently available in your area. Any reference to an IBM product, program, or service is not intended to state or imply that only that IBM product, program, or service may be used. Any functionally equivalent product, program, or service that does not infringe any IBM intellectual property right may be used instead. However, it is the user's responsibility to evaluate and verify the operation of any non-IBM product, program, or service. IBM may have patents or pending patent applications covering subject matter described in this document. The furnishing of this document does not give you any license to these patents. You can send license inquiries, in writing, to: IBM Director of Licensing IBM Corporation North Castle Drive Armonk, NY 10504-1785 U.S.A. For license inquiries regarding double-byte (DBCS) information, contact the IBM Intellectual Property Department in your country or send inquiries, in writing, to: IBM World Trade Asia Corporation Licensing 2-31 Roppongi 3-chome, Minato-ku Tokyo 106, Japan The following paragraph does not apply to the United Kingdom or any other country where such provisions are inconsistent with local law: INTERNATIONAL BUSINESS MACHINES CORPORATION PROVIDES THIS PUBLICATION "AS IS" WITHOUT WARRANTY OF ANY KIND, EITHER EXPRESS OR IMPLIED, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF NON-INFRINGEMENT, MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE. Some states do not allow disclaimer of express or implied warranties in certain transactions, therefore, this statement may not apply to you. This information could include technical inaccuracies or typographical errors. Changes are periodically made to the information herein; these changes will be incorporated in new editions of the publication. IBM may make improvements and/or changes in the product(s) and/or the program(s) described in this publication at any time without notice. Any references in this information to non-IBM websites are provided for convenience only and do not in any manner serve as an endorsement of those websites. The materials at those websites are not part of the materials for this IBM product and use of those websites is at your own risk. IBM may use or distribute any of the information you supply in any way it believes appropriate without incurring any obligation to you. © Copyright IBM Corp. 1990, 2012 311 Licensees of this program who wish to have information about it for the purpose of enabling: (i) the exchange of information between independently created programs and other programs (including this one) and (ii) the mutual use of the information which has been exchanged, should contact: Lab Director IBM Canada Ltd. Laboratory 8200 Warden Avenue Markham, Ontario L6G 1C7 Canada Such information may be available, subject to appropriate terms and conditions, including in some cases, payment of a fee. The licensed program described in this document and all licensed material available for it are provided by IBM under terms of the IBM Customer Agreement, IBM International Program License Agreement or any equivalent agreement between us. Any performance data contained herein was determined in a controlled environment. Therefore, the results obtained in other operating environments may vary significantly. Some measurements may have been made on development-level systems and there is no guarantee that these measurements will be the same on generally available systems. Furthermore, some measurements may have been estimated through extrapolation. Actual results may vary. Users of this document should verify the applicable data for their specific environment. Information concerning non-IBM products was obtained from the suppliers of those products, their published announcements or other publicly available sources. IBM has not tested those products and cannot confirm the accuracy of performance, compatibility or any other claims related to non-IBM products. Questions on the capabilities of non-IBM products should be addressed to the suppliers of those products. All statements regarding IBM's future direction or intent are subject to change or withdrawal without notice, and represent goals and objectives only. This information contains examples of data and reports used in daily business operations. To illustrate them as completely as possible, the examples include the names of individuals, companies, brands, and products. All of these names are fictitious and any similarity to the names and addresses used by an actual business enterprise is entirely coincidental. COPYRIGHT LICENSE: This information contains sample application programs in source language, which illustrates programming techniques on various operating platforms. You may copy, modify, and distribute these sample programs in any form without payment to IBM, for the purposes of developing, using, marketing or distributing application programs conforming to the application programming interface for the operating platform for which the sample programs are written. These examples have not been thoroughly tested under all conditions. IBM, therefore, cannot guarantee or imply reliability, serviceability, or function of these programs. You may copy, modify, and distribute these sample programs in any form without payment to IBM for the purposes of developing, using, marketing, or distributing application programs conforming to IBM's application programming interfaces. 312 XL Fortran: Optimization and Programming Guide Each copy or any portion of these sample programs or any derivative work, must include a copyright notice as follows: © (your company name) (year). Portions of this code are derived from IBM Corp. Sample Programs. © Copyright IBM Corp. 1998, 2010. This software and documentation are based in part on the Fourth Berkeley Software Distribution under license from the Regents of the University of California. We acknowledge the following institution for its role in this product's development: the Electrical Engineering and Computer Sciences Department at the Berkeley campus. Trademarks and service marks IBM, the IBM logo, and ibm.com are trademarks or registered trademarks of International Business Machines Corp., registered in many jurisdictions worldwide. Other product and service names might be trademarks of IBM or other companies. A current list of IBM trademarks is available on the web at “Copyright and trademark information” at http://www.ibm.com/legal/copytrade.shtml. Adobe, the Adobe logo, PostScript, and the PostScript logo are either registered trademarks or trademarks of Adobe Systems Incorporated in the United States, other countries, or both. Linux is a registered trademark of Linus Torvalds in the United States, other countries, or both. Microsoft and Windows are trademarks of Microsoft Corporation in the United States, other countries, or both. Cell Broadband Engine is a trademark of Sony Computer Entertainment, Inc. in the United States, other countries, or both and is used under license therefrom. UNIX is a registered trademark of The Open Group in the United States and other countries. Other company, product, or service names may be trademarks or service marks of others. Notices 313 314 XL Fortran: Optimization and Programming Guide Index assembler Compiler-friendly techniques (continued) Special characters low-level linkage conventions 264 compiler invocations 51 _OPENMP C preprocessor macro 73 ATOMIC 93 floating-point precision 52 -O0 3 PERMUTATION 52 -O2 3 pointers 51 -O3 5 B variables 51 trade-offs 6 WHERE constructs 52 BARRIER 97 -O4 7 computed GOTO basic example, described xii trade-offs 7 code size 47 basic optimization 2 -O5 8 conditional vector merge intrinsic bitwise-identical floating-point trade-offs 8 functions (FORTRAN 77 results 293 -qarch 31 extension) 302 block -qcache 7, 33 control and status register for floating cyclic scheduling 169 -qcompact point 297 blocked special files, interaction of XL code size 47 controlling optimization Fortran I/O with 277 -qfloat compiler option 293 transformations 38 buffers, flushing 278 fltint suboption 293 COPYIN clause 150 BYTE data type (FORTRAN 77 hssngl suboption 293 core file 295, 296 extension) 302 nans suboption 299 CRAY functions (FORTRAN 77 extension) nomaf suboption 293 conditional vector merge rsqrt suboption 293 intrinsics 302 -qflttrap compiler option 294 C date and time service and utility -qfunctrace 25 C language and interlanguage calls 253, functions 302 -qfunctrace_xlf_enter 25 256 CRAY pointer (FORTRAN 77 extension), -qfunctrace_xlf_exit 25 C++ and Fortran in same program 255 XL Fortran equivalent 302 -qfuntrace_xlf_catch 25 calling by reference or value 261 CRITICAL 99 -qinline inlining calling non-Fortran procedures 253 CVMGx intrinsic functions (FORTRAN 77 code size 46 CASE construct extension) 302 -qipa 7 code size 47 IPA process 7 character data, passing between -qlistfmt compiler option 18 languages 259 D -qnofunctrace 25 character special files, interaction of XL data scope attribute clauses -qposition compiler option 276 Fortran I/O with 277 COPYIN clause 150 -qsigtrap compiler option 295 character-count edit descriptor discussion 146 -qstrict 6 (FORTRAN 77 extension) 302 data types in Fortran, C 257 -qtune 32 check_fpscr.f sample file 299 date and time functions (FORTRAN 77 /opt/ibmcmp/xlf/bg/14.1/include/ chunk extension) 302 fexcp.h 295 SCHEDULE directive and 120 debugging 20 /tmp directory 279 code size 45 delays runtime option 77 * length specifiers (FORTRAN 77 -qcompact 47 deprecated SMP directive 93 extension) 302 -qinline inlining 46 Directive Clauses, Global Rules 146 %REF functions 261 array initialization 46 directives 93 %VAL functions 261 blocking 45 DO SERIAL 103 CASE constructs 47 END MASTER 107 computed GOTOs 47 END ORDERED 108 dynamic linking 48 Numerics END PARALLEL 111 high activity areas 47 64-bit data types (FORTRAN 77 END PARALLEL DO 113 page faults 47 extension) 302 END PARALLEL SECTIONS 117 static linking 48 END PARALLEL WORKSHARE 119 steps for reduction 46 END SECTIONS 123 common blocks FLUSH 105 A in data scope attribute clauses 146 MASTER 107 advanced optimization 4 compiler directives nesting 90 aliasing 41 speculative do 130 ORDERED 108 arguments speculative section 131 PARALLEL 111 passing between languages 256, 257 TM_ATOMIC 142 PARALLEL DO 113 passing by reference or by value 261 Compiler-friendly techniques 51 PARALLEL SECTIONS 117 array initialization aliasing 51 PARALLEL WORKSHARE 119 code size 46 arrays 52 SCHEDULE 120 arrays choosing appropriate variable SECTIONS 123 passing between languages 260 sizes 52 SINGLE / END SINGLE 126 © Copyright IBM Corp. 1990, 2012 315 directives (continued) f_pthread_mutexattr_getpshared THREADLOCAL 135 F function 235 THREADPRIVATE 138 f_maketime function 204 f_pthread_mutexattr_gettype WORKSHARE 144 f_pthread 201 function 236 Directives f_pthread_attr_destroy function 205 f_pthread_mutexattr_init function 237 ATOMIC 93 f_pthread_attr_getdetachstate f_pthread_mutexattr_setpshared BARRIER 97 function 205 function 237 CRITICAL 99 f_pthread_attr_getguardsize f_pthread_mutexattr_settype DO (worksharing) 100 function 206 function 238 END CRITICAL 99 f_pthread_attr_getinheritsched f_pthread_mutexattr_t 239 END DO (worksharing) 100 function 206 f_pthread_once function 239 DO (worksharing) 100 f_pthread_attr_getschedparam f_pthread_once_t 239 DO SERIAL compiler directive 103 function 207 f_pthread_rwlock_destroy function 240 double-precision values 286, 288 f_pthread_attr_getschedpolicy f_pthread_rwlock_init function 240 dynamic linking function 208 f_pthread_rwlock_rdlock function 241 code size 48 f_pthread_attr_getscope function 208 f_pthread_rwlock_t function 241 f_pthread_attr_getstack function 209 f_pthread_rwlock_tryrdlock f_pthread_attr_init function 209 function 242 E f_pthread_attr_setdetachstate f_pthread_rwlock_trywrlock function 210 function 242 enable suboption of -qflttrap 296 f_pthread_attr_setguardsize 211 f_pthread_rwlock_unlock function 243 enabling MASS 10 f_pthread_attr_setinheritsched f_pthread_rwlock_wrlock function 243 END CRITICAL 99 function 211 f_pthread_rwlockattr_destroy END DO (worksharing) 100 f_pthread_attr_setschedparam function 244 END MASTER compiler directive 107 function 212 f_pthread_rwlockattr_getpshared END ORDERED compiler directive 108 f_pthread_attr_setschedpolicy function 244 END PARALLEL compiler directive 111 function 213 f_pthread_rwlockattr_init function 245 END PARALLEL DO compiler f_pthread_attr_setscope function 213 f_pthread_rwlockattr_setpshared directive 113 f_pthread_attr_setstack function 214 function 246 END PARALLEL SECTIONS compiler f_pthread_attr_t function 215 f_pthread_rwlockattr_t function 246 directive 117 f_pthread_cancel function 215 f_pthread_self function 247 END PARALLEL WORKSHARE compiler f_pthread_cleanup_pop function 215 f_pthread_setcancelstate function 247 directive 119 f_pthread_cleanup_push function 216 f_pthread_setcanceltype function 248 END SECTIONS compiler directive 123 f_pthread_cond_broadcast function 217 f_pthread_setconcurrency function 249 environment variables f_pthread_cond_destroy function 218 f_pthread_setschedparam function 249 BG_SMP_FAST_WAKEUP 78 f_pthread_cond_init function 218 f_pthread_setspecific function 250 BG_SPEC_SCRUB_CYCLE 79 f_pthread_cond_signal function 219 f_pthread_t function 251 OpenMP f_pthread_cond_t function 219 f_pthread_testcancel function 251 OMP_DYNAMIC 79 f_pthread_cond_timedwait function 220 f_sched_param function 251 OMP_NESTED 80 f_pthread_cond_wait function 220 f_sched_yield function 251 OMP_NUM_THREADS 80 f_pthread_condattr_destroy function 221 f_timespec function 252 OMP_PROC_BIND 82 f_pthread_condattr_getpshared fexcp.h include file 295 OMP_SCHEDULE 82 function 221 fhandler.F sample file 299 OMP_STACKSIZE 83 f_pthread_condattr_init function 222 file positioning 276 OMP_WAIT_POLICY 84 f_pthread_condattr_setpshared files runtime function 223 I/O formats 273 XLSMPOPTS 74 f_pthread_condattr_t function 223 names 274 SE_MAX_NUM_ROLLBACK 85 f_pthread_create function 224 permissions 277 SE_REPORT_ENABLE 85 f_pthread_detach function 225 floating-point SE_REPORT_LOG 86 f_pthread_equal function 226 exceptions 294 SE_REPORT_NAME 85 f_pthread_exit function 226 processing 285 TM_ENABLE_INTERRUPT_ON_ f_pthread_getconcurrency function 227 optimizing 293 CONFLICT 87 f_pthread_getschedparam function 227 floating-point optimization 36 TM_MAX_NUM_ROLLBACK 86 f_pthread_getspecific function 228 floating-point status and control TM_REPORT_LOG 88 f_pthread_join function 229 register 297 TM_REPORT_NAME 87 f_pthread_key_create function 229 FLTTRAP @PROCESS directive 294 TM_REPORT_STAT_ENABLE 87 f_pthread_key_delete function 230 flttrap_handler.c and flttrap_test.f sample TM_SHORT_TRANSACTION_MODE 89 f_pthread_key_t 231 files 299 exception handling 287 f_pthread_kill function 231 FLUSH compiler directive 105 for floating point 294 f_pthread_mutex_destroy function 231 flushing I/O buffers 278 installing an exception handler 295 f_pthread_mutex_init function 232 formats, file 273 execution environment routines f_pthread_mutex_lock function 233 fort.* default file names 275, 279 OpenMP 173 f_pthread_mutex_t 233 FORTRAN 77 extensions, list of common explicit interfaces 262 f_pthread_mutex_trylock function 233 ones 302 extended-precision values 289 f_pthread_mutex_unlock function 234 fpdt.h and fpdc.h include files 290 extensions to FORTRAN 77, list of f_pthread_mutexattr_destroy fpgets and fpsets service and utility common ones 302 function 234 subroutines 297 316 XL Fortran: Optimization and Programming Guide fpscr register 297 links, interaction of XL Fortran I/O omp_get_team_size OpenMP fpstat array 297 with 277 function 181 functions lock routines omp_get_thread_limit OpenMP linkage convention for calls 269 OpenMP 173 function 181 return values 263 long variable names (FORTRAN 77 omp_get_thread_num execution FUNCTRACE_XLF_CATCH 25 extension) 302 environment routine 181 FUNCTRACE_XLF_ENTER 25 long vectorization 10 omp_get_wtick OpenMP timing FUNCTRACE_XLF_EXIT 25 loop optimization 9 routine 183 lower case (FORTRAN 77 extension) 302 omp_get_wtime OpenMP timing routine 183 G omp_in_final execution environment routine 184 get_round_mode procedure 290 M omp_in_parallel execution environment GETENV intrinsic procedure 274 macro, _OPENMP C preprocessor 73 routine 184 main, restriction on use as a Fortran omp_init_lock lock routine 185 name 253 omp_init_nest_lock OpenMP nestable MASS libraries 55 H lock routine 186 scalar functions 56 high-order transformation 9 OMP_NESTED environment variable 80 vector functions 58 HOT 9 OMP_NUM_THREADS environment MASTER compiler directive 107 variable 80 migrating OMP_PROC_BIND environment from other systems 301 variable 82 I minus infinity, representation of 286 OMP_SCHEDULE environment IEEE arithmetic 285 mixing integers and logicals (FORTRAN variable 82 implicitly connected files 275 77 extension) 302 omp_set_dynamic execution environment include files fpdt.h and fpdc.h 290 module procedures, external names routine 187 infinity values 286 corresponding to 253 omp_set_lock lock routine 187 initial file position 276 omp_set_max_active_levels Inlining 42 subroutine 188 input/output 285 N omp_set_nest_lock nestable lock from two languages in the same naming conventions for external routine 189 program 254 names 253 omp_set_nested execution environment redirection 276 NaN values routine 189 XL Fortran implementation and infinities 286 omp_set_num_threads execution details 273 negative infinity, representation of 286 environment routine 190 integer POINTER (FORTRAN 77 NOFUNCTRACE 25 omp_set_schedule routine 191 extension) 302 null-terminated strings, passing to C OMP_STACKSIZE environment interlanguage calls 253, 261 functions 259 variable 83 arrays 260 omp_test_lock lock routine 192 C++ 255 omp_test_nest_lock lock routine 192 character types 259 omp_unset_lock lock routine 193 corresponding data types 257 O omp_unset_nest_lock lock routine 194 input and output 254 omp_destory_nest_lock OpenMP nestable OMP_WAIT_POLICY environment low-level linkage conventions 264 lock routine 175 variable 84 pointers 261 omp_destroy_lock OpenMP lock OpenMP intrinsic procedures routine 174 execution environment routines reset_speculation_mode 197, 201 OMP_DYNAMIC environment description 173 se_get_all_stats 195 variable 79 omp_get_dynamic 176 se_print_stats 196 omp_get_active_level OpenMP omp_get_max_threads 177 tm_get_all_stats 199 function 175 omp_get_nested 178 tm_get_stats 198 omp_get_ancestor_thread_num OpenMP omp_get_num_procs 179 tm_print_all_stats 200 function 176 omp_get_num_threads 179 tm_print_stats 199 omp_get_dynamic execution environment omp_get_thread_num 181 IPA 11 routine 176 omp_in_final 184 levels 13 omp_get_level OpenMP function 177 omp_in_parallel 184 suboptions 13 omp_get_max_active_levels OpenMP omp_set_dynamic 187 with C and C++ 14 function 177 omp_set_nested 189 irrevocable mode 70 omp_get_max_threads execution omp_set_num_threads 190 environment routine 177 function omp_get_nested execution environment omp_get_level 177 routine 178 L omp_get_max_active_levels 177 omp_get_num_procs execution libmass library 56 omp_get_thread_limit 181 environment routine 179 libmassv library 58 get active level function omp_get_num_threads execution library omp_get_active_level 175 environment routine 179 MASS 55 omp_get_schedule OpenMP schedule scalar 56 routine 180 vector 58 Index 317 OpenMP (continued) PARALLEL SECTIONS compiler directive Pthreads Library Module (continued) get ancestor thread number function discussion 117 f_pthread_cond_broadcast omp_get_ancestor_thread_num 176 PARALLEL WORKSHARE compiler function 217 get team size function directive f_pthread_cond_destroy function 218 omp_get_team_size 181 discussion 119 f_pthread_cond_init function 218 lock routines parthreshold runtime option 77 f_pthread_cond_signal function 219 description 173 Pascal language and interlanguage f_pthread_cond_t function 219 omp_destroy_lock 174 calls 253 f_pthread_cond_timedwait omp_init_lock 185 performance analysis 39 function 220 omp_set_lock 187 performance of floating-point f_pthread_cond_wait function 220 omp_test_lock 192 arithmetic 293 f_pthread_condattr_destroy omp_test_nest_lock 192 performance tuning options 76 function 221 omp_unset_lock 193 permissions of files 277 f_pthread_condattr_getpshared omp_unset_nest_lock 194 pipes, interaction of XL Fortran I/O function 221 nestable lock routines with 277 f_pthread_condattr_init function 222 omp_destroy_nest_lock 175 plus infinity, representation of 286 f_pthread_condattr_setpshared omp_init_nest_lock 186 pointers (integer POINTER) (FORTRAN function 223 omp_set_nest_lock 189 77 extension) 302 f_pthread_condattr_t function 223 schedule routines portability 301 f_pthread_create function 224 omp_get_schedule 180 porting to XL Fortran 301 f_pthread_detach function 225 omp_set_schedule 191 POSITION @PROCESS directive 276 f_pthread_equal function 226 subroutines position of a file after an OPEN f_pthread_exit function 226 omp_set_max_active_levels 188 statement 276 f_pthread_getconcurrency timing routines positive infinity, representation of 286 function 227 omp_get_wtick 183 postmortem.f sample file 299 f_pthread_getschedparam omp_get_wtime 183 preconnected files 275 function 227 OpenMP environment variables 79, 83, procedures 25 f_pthread_getspecific function 228 84 profilefreq runtime option 78 f_pthread_join function 229 optimization providing your application f_pthread_key_create function 229 -O0 3 characteristics 36 f_pthread_key_delete function 230 -O2 3 pseudo-devices, interaction of XL Fortran f_pthread_key_t 231 -O3 5 I/O with 277 f_pthread_kill function 231 -O4 7 pthreads library module 309 f_pthread_mutex_destroy -O5 8 Pthreads Library Module function 231 advanced 4 descriptions of functions in 201 f_pthread_mutex_init function 232 and tuning 31 f_maketime function 204 f_pthread_mutex_lock function 233 basic 2 f_pthread_attr_destroy function 205 f_pthread_mutex_t 233 debugging 20 f_pthread_attr_getdetachstate f_pthread_mutex_trylock floating-point 36, 38 function 205 function 233 for floating-point arithmetic 293 f_pthread_attr_getguardsize f_pthread_mutex_unlock loops 36, 38 function 206 function 234 math functions 55 f_pthread_attr_getinheritsched f_pthread_mutexattr_destroy options to avoid 40 function 206 function 234 optimization and tuning f_pthread_attr_getschedparam f_pthread_mutexattr_getpshared optimizing 1 function 207 function 235 tuning 1 f_pthread_attr_getschedpolicy f_pthread_mutexattr_gettype optimization trade-offs function 208 function 236 -O3 6 f_pthread_attr_getscope function 208 f_pthread_mutexattr_init -O4 7 f_pthread_attr_getstack 209 function 237 -O5 8 f_pthread_attr_init function 209 f_pthread_mutexattr_setpshared optimization, diagnostics 18, 20 f_pthread_attr_setdetachstate function 237 optimizing function 210 f_pthread_mutexattr_settype applications 1 f_pthread_attr_setguardsize function 238 techniques 8 function 211 f_pthread_mutexattr_t 239 option driven tuning 35 f_pthread_attr_setinheritsched f_pthread_once function 239 OPTIONAL attribute 263 function 211 f_pthread_once_t 239 options for targeting your f_pthread_attr_setschedparam f_pthread_rwlock_destroy architecture 31 function 212 function 240 ORDERED compiler directive 108 f_pthread_attr_setschedpolicy f_pthread_rwlock_init function 240 function 213 f_pthread_rwlock_rdlock f_pthread_attr_setscope function 213 function 241 P f_pthread_attr_setstack function 214 f_pthread_rwlock_t function 241 f_pthread_attr_t function 215 f_pthread_rwlock_tryrdlock PARALLEL compiler directive f_pthread_cancel function 215 function 242 discussion 111 f_pthread_cleanup_pop function 215 f_pthread_rwlock_trywrlock PARALLEL DO compiler directive f_pthread_cleanup_push function 242 discussion 113 function 216 SCHEDULE clause 113 318 XL Fortran: Optimization and Programming Guide Pthreads Library Module (continued) scratch_vars runtime option 279 tuning (continued) f_pthread_rwlock_unlock SECTIONS compiler directive providing your application function 243 discussion 123 characteristics 36 f_pthread_rwlock_wrlock seqthreshold runtime option 78 typeless constants (FORTRAN 77 function 243 short vectorization 10 extension) 302 f_pthread_rwlockattr_destroy SIGFPE signal 294, 295 function 244 signal handling f_pthread_rwlockattr_getpshared for floating point 294 U function 244 installing an exception handler 295 unit_vars runtime option 279 f_pthread_rwlockattr_init signaling NaN 286, 299 usrthds runtime option 76 function 245 SIGTRAP signal 294, 295 f_pthread_rwlockattr_setpshared SIMD 33 function 246 SINGLE / END SINGLE compiler f_pthread_rwlockattr_t function 246 directive 126 V f_pthread_self function 247 single-precision values 286, 288 VALUE attribute 261 f_pthread_setcancelstate function 247 SMP value, passing arguments by 261 f_pthread_setcanceltype function 248 sample programs 305 vector MASS library 58 f_pthread_setchedparam SMP applications vectorization function 249 Developing and running 89 long 10 f_pthread_setconcurrency special files, interaction of XL Fortran I/O short 10 function 249 with 277 f_pthread_setspecific function 250 spins runtime option 77 f_pthread_t function 251 stack 265 W f_pthread_testcancel function 251 stack runtime option 76 work-sharing constructs f_sched_param function 251 standard error, input, and output SECTIONS / END SECTIONS f_sched_yield function 251 streams 275 compiler directives 123 f_timespec function 252 star length specifiers 302 SINGLE / END SINGLE compiler Pthreads Library, BlueGene 201 static linking directives 126 Pthreads Library, Linux 201 code size 48 WORKSHARE compiler directive 144 status and control register for floating point 297 Q stderr, stdin, and stdout streams 275 strings, passing to C functions 259 X Q (character-count) edit descriptor subprograms in other languages, xl__ieee.F and xl__ieee.c sample (FORTRAN 77 extension) 302 calling 253, 256 files 299 quiet NaN 286 Sun pointer (FORTRAN 77 extension), XL xl__sigdump exception handler 295 Fortran equivalent 302 xl__trbk exception handler 295 supervised mode 70 xl__trbk_test.f sample file 299 R symbolic links, interaction of XL Fortran xl__trce exception handler 295 real arithmetic 285 I/O with 277 xl__trcedump exception handler 295 REAL(16) values 289 system Pthreads Library 201 XLFSCRATCH_unit environment REAL(4) and REAL(8) values 286, 288 variable 279 record lengths 277 XLFUNIT_unit environment redirecting input/output 276 T variable 279 reference, passing arguments by 261 XLSMPOPTS environment variable 74 tctl command 277 rounding 290 XML report schema 20 thread visible variables 105 rounding errors 292 thread-safing rounding mode 290, 292 pthreads library module 201 THREADLOCAL compiler directive 135 Y THREADPRIVATE compiler yields runtime option 77 S directive 138 safe mode 70 time and date functions (FORTRAN 77 sample programs 305 extension) 302 calling C functions from Fortran 257 Trace/BPT trap 295 floating-point exception traceback listing 295 handling 299 tracing 25 for SMP 305 transactional memory 70 scalar MASS library 56 tuning 31 SCHEDULE clause, of PARALLEL DO -qarch 31 directive 113 -qcache 33 SCHEDULE compiler directive 120 -qtune 32 discussion 120 controlling optimization schedule runtime option 74 transformations 38 scheduling, block cyclic 169 option driven tuning 35 scope options to avoid 40 data scope attribute clauses 146 performance analysis 39 scratch file directory 279 Index 319 320 XL Fortran: Optimization and Programming Guide Product Number: 5799-AH1 Printed in USA SC14-7370-00